docs: known-issues — cite PRs #27, #28, #29

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

.gitea/workflows/ci.yml

@@ -9,32 +9,89 @@ jobs:
     runs-on: ubuntu-latest
     container: rust:latest
     steps:
-      - uses: actions/checkout@v4
-      - name: Cache cargo registry/target
-        uses: actions/cache@v4
-        with:
-          path: |
-            ~/.cargo/registry
-            ~/.cargo/git
-            target
-          key: ${{ runner.os }}-cargo-${{ hashFiles('**/Cargo.lock') }}
+      # Plain git rather than actions/checkout: that is a JavaScript action,
+      # and rust:latest has no `node`, so it failed with exit 127 before any
+      # code was built — on every push. actions/cache went for the same reason.
+      - name: Check out
+        run: |
+          git init -q .
+          git remote add origin "${GITHUB_SERVER_URL}/${GITHUB_REPOSITORY}.git"
+          for i in 1 2 3; do git fetch -q --depth 1 origin "${GITHUB_SHA}" && break; sleep 5; done
+          git checkout -q FETCH_HEAD
       - name: Install rustfmt & clippy components
         run: rustup component add rustfmt clippy
       - name: Install thumbv7em-none-eabihf target
         run: rustup target add thumbv7em-none-eabihf
+      - name: Install wasm32-unknown-unknown target
+        # ci-test.sh builds the reader and clawhdf5-wasm for the browser.
+        run: rustup target add wasm32-unknown-unknown
       - name: Install Python interop dependencies
         # The interop suites used to skip silently when python3/h5py were
         # missing, so they never ran in CI. Install them and make a missing
         # dependency a failure (CLAWHDF5_REQUIRE_INTEROP below).
         run: |
           apt-get update
-          apt-get install -y --no-install-recommends python3 python3-venv
+          # cmake builds libz-ng-sys for the opt-in `fast-deflate` (zlib-ng)
+          # steps in ci-test.sh; rust:latest does not ship it. The default
+          # build (pure-Rust zlib-rs) does not need it.
+          # hdf5-tools: h5ls/h5stat/h5dump/h5diff, which the h5rs
+          # (clawhdf5-tools) interop tests compare against.
+          apt-get install -y --no-install-recommends python3 python3-venv cmake hdf5-tools
           python3 -m venv /opt/interop
-          /opt/interop/bin/pip install --no-cache-dir h5py numpy netCDF4 xarray
+          # maturin + pytest: ci-test.sh builds the Python package
+          # (crates/clawhdf5-py) and runs its tests against h5py.
+          /opt/interop/bin/pip install --no-cache-dir h5py numpy netCDF4 xarray h5netcdf hdf5plugin maturin pytest
           echo "/opt/interop/bin" >> "$GITHUB_PATH"
       - name: Show interop library versions
-        run: python3 -c "import h5py, netCDF4; print('h5py', h5py.__version__, 'HDF5', h5py.version.hdf5_version, 'netCDF4', netCDF4.__version__)"
+        # h5dump's version too: the h5rs dump test requires its exact output
+        # (checked against Debian's 1.14.5 in rust:latest and 1.14.6).
+        run: |
+          /opt/interop/bin/python -c "import h5py, netCDF4, hdf5plugin; print('h5py', h5py.__version__, 'HDF5', h5py.version.hdf5_version, 'netCDF4', netCDF4.__version__, 'hdf5plugin', hdf5plugin.version)"
+          h5dump --version
       - name: Run CI script
         env:
+          # Name the interpreter outright rather than relying on $GITHUB_PATH
+          # reaching the test processes: if `python3` resolved to the system
+          # one instead of the venv, every interop suite would skip.
+          # CLAWHDF5_REQUIRE_INTEROP turns that skip into a failure, so the
+          # two together mean the suites either run or the build goes red.
+          CLAWHDF5_PYTHON: /opt/interop/bin/python
           CLAWHDF5_REQUIRE_INTEROP: "1"
         run: bash scripts/ci-test.sh
+
+  test-arm64:
+    # The aarch64 kernels in clawhdf5-accel — NEON `dot_i8`, including the
+    # SDOT path, and the f32 NEON kernels — are cfg'd out on x86, so the job
+    # above never compiles, lints or tests them.
+    #
+    # `linux_arm64` is served by two runners that execute differently:
+    # vision-01 runs steps on the host (Rust already installed) and vision-02
+    # runs them in docker.gitea.com/runner-images. So the steps work in both:
+    # no `container:`, no JavaScript actions (they are fetched from GitHub,
+    # which not every runner reliably reaches), and an explicit `+stable`
+    # toolchain rather than whatever a host happens to default to.
+    runs-on: linux_arm64
+    env:
+      CARGO_NET_RETRY: "10"
+      CARGO_TERM_COLOR: always
+    steps:
+      - name: Check out
+        run: |
+          git init -q .
+          git remote add origin "${GITHUB_SERVER_URL}/${GITHUB_REPOSITORY}.git"
+          for i in 1 2 3; do git fetch -q --depth 1 origin "${GITHUB_SHA}" && break; sleep 5; done
+          git checkout -q FETCH_HEAD
+      - name: Rust stable
+        run: |
+          export PATH="$HOME/.cargo/bin:$PATH"
+          command -v rustup >/dev/null || curl -sSf --retry 5 https://sh.rustup.rs | sh -s -- -y --profile minimal --default-toolchain none
+          rustup toolchain install stable --profile minimal --component clippy
+          echo "$HOME/.cargo/bin" >> "$GITHUB_PATH"
+      - name: Confirm aarch64
+        run: |
+          test "$(uname -m)" = aarch64
+          if grep -q asimddp /proc/cpuinfo; then echo "dot-product extension present: SDOT kernel runs"; else echo "no dot-product extension: plain NEON kernel runs"; fi
+      - name: Clippy (aarch64 kernels)
+        run: cargo +stable clippy -p clawhdf5-accel --all-targets -- -D warnings
+      - name: Test
+        run: cargo +stable test -p clawhdf5-accel -p clawhdf5-ann -p clawhdf5-format

.gitea/workflows/conformance.yml

@@ -0,0 +1,58 @@
+name: Conformance
+# Nightly: read every file of the pinned public HDF5 corpora with clawhdf5 and
+# with h5py/libhdf5 and compare (conformance/run.sh; CONFORMANCE.md explains
+# the method). Fails on any panic, hang, crash or out-of-memory in clawhdf5,
+# and when the ok count drops below conformance/baseline.json or a file the
+# baseline lists as ok stops being ok. The report is printed into the job log;
+# nothing is uploaded (artifact actions are JavaScript, which rust:latest
+# cannot run — see CLAUDE.md).
+on:
+  schedule:
+    - cron: "17 3 * * *"
+  workflow_dispatch:
+jobs:
+  conformance:
+    runs-on: ubuntu-latest
+    container: rust:latest
+    timeout-minutes: 60
+    env:
+      CARGO_NET_RETRY: "10"
+    steps:
+      # Plain git, not actions/checkout (a JavaScript action; see ci.yml).
+      - name: Check out
+        run: |
+          git init -q .
+          git remote add origin "${GITHUB_SERVER_URL}/${GITHUB_REPOSITORY}.git"
+          for i in 1 2 3; do git fetch -q --depth 1 origin "${GITHUB_SHA}" && break; sleep 5; done
+          git checkout -q FETCH_HEAD
+      - name: Install h5py, h5dump and the probe's codec libraries
+        # hdf5-tools: h5dump for the CVE-corpus comparison. libaec-dev and
+        # pkg-config: the probe builds clawhdf5-format with `szip` (the core
+        # crates' default build needs neither).
+        run: |
+          apt-get update
+          apt-get install -y --no-install-recommends python3 python3-venv hdf5-tools libaec-dev pkg-config
+          python3 -m venv /opt/conformance
+          /opt/conformance/bin/pip install --no-cache-dir -r conformance/requirements.txt
+          /opt/conformance/bin/python -c "import h5py, hdf5plugin; print('h5py', h5py.__version__, 'HDF5', h5py.version.hdf5_version, 'hdf5plugin', hdf5plugin.version)"
+          h5dump --version
+      - name: Probe unit tests
+        run: cargo test --release --manifest-path conformance/probe/Cargo.toml
+        env:
+          CARGO_TARGET_DIR: conformance/.cache/target
+      - name: Reference-side tests
+        run: /opt/conformance/bin/python conformance/test_ref.py
+      - name: Sweep
+        # The corpora come from GitHub (pinned commits, conformance/corpus.txt),
+        # so this job needs a runner that reaches github.com.
+        env:
+          CLAWHDF5_PYTHON: /opt/conformance/bin/python
+        run: bash conformance/run.sh
+      - name: Report
+        if: always()
+        run: |
+          if [ -f CONFORMANCE.md ]; then cat CONFORMANCE.md; else echo "no report was generated"; fi
+          if [ -f conformance/.cache/results/summary.md ]; then
+            echo; echo "---- per-file detail (conformance/.cache/results/summary.md) ----"
+            cat conformance/.cache/results/summary.md
+          fi

.gitignore

@@ -4,3 +4,9 @@ benchmarks/longmemeval/*.json
 
 # Local model weights (MiniLM etc.) — large, not committed
 weights/
+.venv
+__pycache__/
+.pytest_cache/
+
+# Scratch files the heavy tests generate (huge_chunks_interop)
+crates/*/tests/scratch/

CLAUDE.md

@@ -1,120 +1,275 @@
 # clawhdf5
 
 ## Purpose
-Pure-Rust HDF5 format implementation with HNSW vector search, WAL-backed persistence, agent memory storage, and GPU-accelerated I/O. Used by ZeroClaw as its persistent memory and knowledge graph backend.
+Pure-Rust HDF5 implementation (read, write, in-place edit, remote and browser
+reads) plus agent memory on top of it: HNSW vector search, a WAL-backed store,
+and GPU vector distances. A standalone library. Its one verified consumer is
+ClawBrainHub (`.brain` files); no agent framework integrates it (see
+*Standing rules*).
 
 ## Architecture
 
-Cargo workspace with 16 crates under `crates/` (plus `libaec-sys`, an internal FFI bindings crate for the optional `szip` feature):
+Cargo workspace, 19 crates under `crates/` (plus `libaec-sys`, the FFI crate
+behind the optional `szip` feature). MSRV 1.92 (`rust-version`, checked in CI).
 
 | Crate | Role |
 |-------|------|
-| `clawhdf5-format` | HDF5 binary spec parser (superblock, B-tree, heap) — also holds shared type definitions and physical constants |
-| `clawhdf5-io` | Read/write implementation |
-| `clawhdf5-filters` | Compression filters (gzip, LZ4, Zstd, Blosc) |
-| `clawhdf5-derive` | Proc-macro derive for HDF5-serializable structs |
-| `clawhdf5` | Main facade crate |
-| `clawhdf5-netcdf4` | NetCDF-4 compatibility layer |
-| `clawhdf5-ann` | HNSW approximate nearest-neighbor vector index |
-| `clawhdf5-agent` | Agent memory, session history, knowledge graph storage |
-| `clawhdf5-gpu` | GPU-accelerated I/O via wgpu (hand-written WGSL compute shaders) |
-| `clawhdf5-accel` | CPU SIMD acceleration path |
-| `clawhdf5-migrate` | Schema migration engine |
+| `clawhdf5-format` | The HDF5 format: parsers and writer (superblock, headers, B-trees, heaps, chunk indexes), the `Storage` trait, the filter pipeline and registry (`filter_registry`), every codec except deflate (LZ4, Zstd, SZIP, N-Bit, scale-offset, pcodec; pure-Rust LZF, bitshuffle, bzip2, Blosc 1; Blosc2 and ZFP read-only), `float16`, `checksum` |
+| `clawhdf5-filters` | Deflate backends (zlib-rs default, zlib-ng, Apple Compression) |
+| `clawhdf5-io` | I/O adapters (buffers, mmap, prefetch) |
+| `clawhdf5-derive` | `#[derive(H5Type)]` for compound types |
+| `clawhdf5` | Facade: `File`, `FileBuilder`, `Dataset`, `FileEditor` (`src/edit/`), SWMR reading (`src/swmr.rs`) |
+| `clawhdf5-netcdf4` | NetCDF-4 read support |
+| `clawhdf5-remote` | `open_url`: HTTP(S) range requests and object stores (S3, GCS, Azure) through `BlockCache` |
+| `clawhdf5-tools` | `h5rs`: `ls`, `dump` (DDL / hdf5-json), `stat`, `diff`, `check` |
+| `clawhdf5-py` | PyO3 bindings (h5py-like API, remote files, `'r+'` editing) |
+| `clawhdf5-wasm` | wasm-bindgen browser reader (`open(bytes)`, `openUrl(url)`); demo in `examples/wasm-viewer/` |
+| `clawhdf5-ann` | HNSW index |
+| `clawhdf5-agent` | Agent memory store (`HDF5Memory`), sessions, knowledge graph, BM25 |
+| `clawhdf5-accel` | CPU SIMD kernels (AVX2, NEON) |
+| `clawhdf5-gpu` | wgpu vector distances (WGSL) — not dataset I/O; HDF5 I/O is CPU-only |
+| `clawhdf5-migrate` | SQLite → agent store migration |
+| `clawhdf5-cli` | Agent-memory CLI |
+| `clawhdf5-napi` | Node.js addon (the `packages/clawhdf5-node` wrapper is broken; `docs/known-issues.md`) |
 | `clawhdf5-android` | Android JNI bindings |
-| `clawhdf5-cli` | Command-line interface |
-| `clawhdf5-napi` | Node.js native addon bindings |
-| `clawhdf5-py` | PyO3 Python bindings |
-| `clawhdf5-bench` | Benchmark suite |
+| `clawhdf5-bench` | Benchmarks and harnesses (`search_harness`, `read_harness`, `concurrent_read`, `longmemeval_bench`, …) |
 
-## Key Features
-- Zero-dependency HDF5 read/write (no libhdf5 C library required)
-- HNSW vector index for semantic similarity search over agent memories — the
-  `clawhdf5-agent` `hnsw` feature is **on by default**, so `hybrid_search` uses
-  the approximate `clawhdf5-ann` index for the vector stage (the index mirrors
-  the cache and self-heals on drift). Build the agent with
-  `--no-default-features --features float16` to force the exact linear cosine scan.
-  The agent's `parallel` feature (also default) builds the index on a thread
-  pool; the graph is identical with or without it.
-  The index uses the HNSW paper's diversity heuristic for neighbour selection
-  (plain closest-M capped recall on clustered data: 0.31 recall@10 at 100K). Its
-  graph is saved to `<store>.h5.ann` at each checkpoint and reloaded by `open()`
-  (tied to the checkpoint by a generation id; stale/damaged sidecars are
-  ignored and the index rebuilt). `hybrid_search` keeps one incremental BM25
-  index for the life of the store and never writes the store: Hebbian
-  activation boosts are persisted by the next checkpoint (or on drop), not per
-  query. Measure any search-path change with
-  `cargo run --release -p clawhdf5-bench --bin search_harness` (baselines in
-  `BENCHMARKS.md`).
-- WAL (write-ahead log) for crash-safe persistence, with a chained CRC32
-  trailer per entry (each entry's CRC folds in the previous entry's CRC) so a
-  corrupted, reordered, duplicated, or spliced entry stops replay cleanly
-  instead of loading bad or tampered data. The pre-chaining per-entry-CRC
-  format (v2) is still fully readable; the oldest no-CRC format (v1) is only
-  reachable through the one-time migration path in `HDF5Memory::open`, not
-  through the public `WalFile::read_entries`.
-  **What the WAL guarantees:** integrity, ordering, and recovery from a
-  *process* crash at any point — including between a checkpoint and the WAL
-  truncate (each checkpoint records a `WalMark` in `/meta`, and `open()` skips
-  the WAL prefix the `.h5` already contains, so entries are never applied
-  twice). Checkpoints and snapshots are made durable as a unit (temp file
-  synced, renamed, directory synced). **What it does not guarantee:**
-  individual WAL appends are *not* fsynced (a deliberate latency trade-off), so
-  saves made since the last checkpoint can be lost on power failure or kernel
-  panic. Current header version is 4 (adds the `Update` record used by
-  `save_or_update`); v3 files are read and upgraded in place.
-- A store has a **single writer**: `HDF5Memory::create`/`open` hold an exclusive
-  advisory lock on `<store>.h5.lock` and a second opener gets
-  `MemoryError::Locked`. Use `HDF5Memory::open_read_only` for a lock-free,
-  never-writing point-in-time view (the CLI's `recall`/`stats`/`agents-md`/
-  `export` do). An unreadable WAL (torn header, bad magic) is quarantined to
-  `<store>.h5.wal.corrupt-<ts>` rather than blocking `open()`; a WAL with an
-  unknown *newer* version still fails and is left untouched.
-- `MemoryConfig::compression` uses deflate by default; enable the agent's
-  `zstd` feature to compress embeddings with Zstd instead (links libzstd).
-- `Dataset::verify_provenance()` (clawhdf5 facade, `provenance` feature, on by
-  default) recomputes a dataset's SHA-256 and compares it against the
-  `_provenance_sha256` attribute written automatically on save when
-  `DatasetBuilder::with_provenance` is used. It's opt-in per call, not run
-  automatically on open — it decodes and hashes the whole dataset. The hash
-  is unkeyed (tamper-*evident*, not tamper-*proof*): it detects accidental
-  corruption, not a deliberate actor able to modify both the data and the
-  stored hash.
-- `clawhdf5-agent`'s `HDF5Memory::save`/`save_batch`/`save_or_update` run every
-  write through an in-memory (session-scoped, not persisted to disk)
-  provenance ledger and write-anomaly detector: a content hash per record
-  (`provenance.rs`) for detecting accidental mid-session corruption, plus
-  rate-limit/injection-pattern/source-distribution checks (`anomaly.rs`).
-  Alerts never block a save — drain them with `HDF5Memory::take_anomaly_alerts`.
-  `MemorySource` for this bookkeeping is inferred from the caller-supplied
-  `source_channel` string (a heuristic, not an authenticated trust boundary).
-- GPU-accelerated batch I/O for large dataset processing
-- Python and Node.js bindings for cross-language use
-- NetCDF-4 compatibility for scientific data interop
+Reference docs: `docs/known-issues.md` (open issues table first — check it
+before calling something a bug or a feature), `BENCHMARKS.md` (headline
+numbers first), `CONFORMANCE.md` (generated), `docs/design/range-reads.md`
+and `docs/design/swmr.md`, `CHANGELOG.md` (full detail of every fix).
+
+## Standing rules
+
+- **No C in the default build.** No libhdf5; deflate defaults to pure-Rust
+  zlib-rs (`fast-deflate` opts into zlib-ng, which needs cmake). `ci-test.sh`
+  fails if a C-building crate enters the core crates' default tree. Zstd,
+  SZIP, `https` (ring) and `s3`/`gcs`/`azure` (aws-lc-rs) are opt-in. flate2
+  must keep `runtime_detection` with zlib-rs — without it zlib-rs loses SIMD
+  and inflates 3.5x slower.
+- **Every file we write must open in h5py/libhdf5.** Interop tests compare
+  against h5py and h5dump; `f32` and empty datasets did not open until
+  2026-09-23.
+- **float16 has one implementation:** `clawhdf5_format::float16`.
+- **Claims need evidence.** Performance and integration claims in docs must
+  be measured, dated (with machine and command), or withdrawn. Benchmark
+  numbers are dated records: never edit a measured value, add a new dated
+  section and mark the old one superseded.
+- **OpenClaw is not supported** (decided 2026-09-25): clawhdf5 is not and
+  never was an OpenClaw memory plugin; the old `memory.backend = "clawhdf5"`
+  config was never valid. `docs/openclaw.md` records what a real plugin would
+  need. The `openclaw` module's `ClawhdfBackend` is just `search` with
+  re-rank + confidence on.
+- **ZeroClaw does not use clawhdf5** (checked 2026-09-25 against upstream
+  v0.8.5 and the `osobh/zeroclaw` fork and their history): its memory
+  backends are its own; `clawhdf5-migrate`'s default SQLite layout is not
+  ZeroClaw's schema. Don't reintroduce integration claims without an
+  integration and a test against the real consumer.
+- **known-issues.md:** one entry per bug; when fixed, record it in
+  `CHANGELOG.md` and move the entry to *Fixed (history)* with date, PR,
+  affected releases and what users must do — never delete it.
+
+## HDF5 library: invariants and gotchas
+
+- **Remote/range reads** (`docs/design/range-reads.md`, M0-M5 merged in PRs
+  #17-#19, M4 listing costs cut in #21): every format-crate read path goes through `Storage`
+  (`read_at`/`read_ranges`/`hint`). `File::open_storage` takes any
+  `Storage`; `clawhdf5_remote::open_url` wraps HTTP (`HttpStorage`, ureq) or
+  `ObjectStoreStorage` in `BlockCache` (1 MiB blocks, LRU budget, in-flight
+  dedup, coalesced runs). Remote files are pinned by ETag/Last-Modified and
+  length (`RemoteError::FileChanged`). Zero-copy APIs and `File::as_bytes`
+  need an in-memory file. Parse through `File::storage()` and the `*_in`
+  functions, not `as_bytes`, in new code (the Python bindings do).
+  `ObjectStoreStorage` runs reads on its own small tokio runtime, so it
+  works from any thread.
+- **SWMR** (`docs/design/swmr.md`): `File::open_swmr` reads a file a libhdf5
+  SWMR writer is appending to — positioned reads, no chunk cache, bounded
+  retries (100), `Dataset::refresh()`. clawhdf5 has no SWMR writer; remote
+  SWMR is out of scope.
+- **Browser** (`clawhdf5-wasm`, read-only, no Zstd/SZIP): `openUrl` reads
+  through the restartable "NeedBytes" cache (`src/lazy.rs`: a call is re-run
+  after each wave of misses; no block is evicted while a call runs); the HTTP
+  is JavaScript (`js/remote.js`).
+- **In-place editing** (`clawhdf5::FileEditor`): overwrites values, grows and
+  shrinks chunked datasets (every chunk index) and sets attributes (compact
+  and dense) without rewriting the file, changing indexes and heaps as
+  libhdf5 does; freed space is reused within one editor. Anything it cannot
+  do safely is `Error::Unsupported` before any write (limits in
+  `docs/known-issues.md`). The algorithms follow libhdf5 `hdf5_1_14_6`
+  (github.com/HDFGroup/hdf5). Test changes with `cargo test -p
+  clawhdf5-tools --test edit_interop --test edit_coverage_interop`.
+- **Provenance:** `Dataset::verify_provenance()` (facade `provenance`
+  feature, default) re-hashes a dataset against its `_provenance_sha256`
+  attribute (`DatasetBuilder::with_provenance`). Opt-in per call; unkeyed
+  hash — tamper-evident, not tamper-proof.
+
+## Agent memory: invariants and gotchas
+
+- **Search.** `HDF5Memory::search(query_emb, text, &SearchOptions)` is the
+  full path: optional source-channel filter (before ranking; exact scan of
+  the allowed records when cheaper than `pool × M` index distance
+  evaluations, and as the fallback when the pool comes back short), fusion,
+  activation scaling, optional re-ranking and confidence rejection.
+  `hybrid_search`/`hybrid_search_with` are thin wrappers. It keeps one
+  incremental BM25 index for the life of the store and never writes the
+  store: Hebbian activation boosts are persisted by the next checkpoint (or
+  on drop).
+- **HNSW** (`hnsw` feature, default): the approximate `clawhdf5-ann` index
+  mirrors the cache and self-heals on drift; build the agent with
+  `--no-default-features --features float16` for the exact linear scan.
+  `parallel` (default) builds it on a thread pool with an identical graph.
+  Neighbour selection uses the HNSW paper's diversity heuristic (closest-M
+  capped recall at 0.31 recall@10 at 100K on clustered data). The graph is
+  saved to `<store>.h5.ann` at each checkpoint, tied to it by a generation
+  id; a stale or damaged sidecar is ignored and the index rebuilt.
+- **`MemoryConfig::quantized_index`** (default on for new stores, persisted;
+  older stores load as `false` — guarded by `tests/fixtures/store_v2_5_0.h5`;
+  CLI `create --f32-index`): the index's copy of the embeddings is `i8`, and
+  the query path re-scores candidates against the exact embeddings. The
+  aarch64 kernels (`clawhdf5_accel::dot_i8`, NEON `SDOT` via inline asm) are
+  `cfg`'d out on x86, so x86 CI never compiles them — test on real ARM
+  (`rpivision02`, 10.0.2.3, a Pi 5) or rely on the `test-arm64` job.
+- **`MemoryConfig::float16`** (default on for new stores, persisted; older
+  stores keep `false` — guarded in `tests/float16_store.rs`; CLI `create
+  --f32`): `/memory/embeddings` is IEEE half. `MemoryCache::half_precision`
+  rounds each embedding as it enters the cache (push, update, WAL replay, and
+  load of a store still `f32` on disk) so memory and file agree bit for bit.
+  Values beyond ±65504 are `MemoryError::InvalidEntry`. The agent's
+  `h5py_interop` test guards that a whole store opens in h5py.
+- **WAL.** Chained CRC32 per entry (a corrupted, reordered, duplicated or
+  spliced entry stops replay cleanly). Header version 4 (`Update` record for
+  `save_or_update`); v3 is upgraded in place, v2 read, v1 only through the
+  one-time migration in `HDF5Memory::open`. Each checkpoint records a
+  `WalMark` in `/meta` so `open()` never applies an entry twice; checkpoints
+  and snapshots are durable as a unit (temp file synced, renamed, directory
+  synced). Individual WAL appends are **not** fsynced (deliberate): saves
+  since the last checkpoint can be lost on power failure or kernel panic.
+- **Single writer.** `create`/`open` hold an exclusive lock on
+  `<store>.h5.lock` (`MemoryError::Locked` for a second opener);
+  `open_read_only` is a lock-free point-in-time view (CLI `recall`/`stats`/
+  `agents-md`/`export`). An unreadable WAL is quarantined to
+  `<store>.h5.wal.corrupt-<ts>`; a WAL of an unknown newer version fails and
+  is left untouched.
+- **Signed checkpoints** (`signing` module): with `set_signing_key` each
+  checkpoint stores an Ed25519-signed manifest (per-record SHA-256 in a
+  Merkle tree plus settings/sessions/graph hashes; `/integrity/record_hashes`);
+  `HDF5Memory::verify(path, &pk)` locates edits. The hashes must cover
+  exactly what the file persists in the form the loader returns it (strings
+  lose trailing NULs; an empty WAL mark is not written) —
+  `tests/signed_store.rs` round-trips awkward strings. The key is never
+  persisted; a signed store refuses to checkpoint without it
+  (`MemoryError::SigningKeyRequired`; `MemoryError` is `#[non_exhaustive]`).
+  WAL entries after the checkpoint are not covered.
+- **Write bookkeeping.** `save`/`save_batch`/`save_or_update` feed an
+  in-memory, session-scoped provenance ledger and anomaly detector
+  (`provenance.rs`, `anomaly.rs`); alerts never block a save
+  (`take_anomaly_alerts`). `MemorySource` is inferred from the caller's
+  `source_channel` string — a heuristic, not a trust boundary.
+- `MemoryConfig::compression` is off by default (deflate, or Zstd with the
+  agent's `zstd` feature, which links libzstd).
 
 ## Workflows
 
-### Build
+Put `$HOME/.cargo/bin` on `PATH`. The h5py/netCDF4 interop tests find their
+Python through `CLAWHDF5_PYTHON` (or `.venv/bin/python`); create it with
+`python3 -m venv .venv && .venv/bin/pip install h5py numpy netCDF4 hdf5plugin`.
+Set `CLAWHDF5_REQUIRE_INTEROP=1` to make a missing interpreter a failure.
+
 ```bash
 cargo build --release
+cargo test --workspace
+bash scripts/ci-test.sh          # everything CI runs (see below)
 ```
 
-### Test
+### CI (`.gitea/workflows/`)
+- **`ci.yml` `test`** (`ubuntu-latest`, `rust:latest` container; runners
+  `tank`, `architect`): installs h5py/netCDF4/xarray/hdf5plugin/maturin/pytest,
+  `hdf5-tools` and `cmake`, then runs `scripts/ci-test.sh` with
+  `CLAWHDF5_REQUIRE_INTEROP=1`. The script runs: fmt; clippy (workspace, the
+  format feature matrix, each plugin filter alone, parallel, fast-deflate,
+  remote with all backends, h5rs remote); "no C in the default build";
+  wasm32 build and clippy; `check-32bit-casts.sh`; the wasm package under
+  Node when `node` and `wasm-bindgen` exist (not in CI); the MSRV check;
+  `cargo test` (workspace plus feature variants: format matrix, parallel,
+  remote/object_store, h5rs URLs, ann parallel, fast-deflate); the h5py
+  interop suites (`writer_h5py_tests --include-ignored`, plugin filters,
+  ZFP); the Python package (clippy, `maturin build`, pytest vs h5py);
+  `cargo bench --no-run`; `check-nostd.sh`; an optional fuzz smoke run
+  (`CLAWHDF5_FUZZ_SECONDS`).
+- **`ci.yml` `test-arm64`** (`linux_arm64`; `vision-01` host mode,
+  `vision-02` Docker — steps must work in both): clippy of
+  `clawhdf5-accel`, tests of `-accel`, `-ann`, `-format`; the only place the NEON kernels build.
+- **`conformance.yml`** (nightly 03:17 UTC and manual): probe unit tests,
+  `conformance/test_ref.py`, then `conformance/run.sh` (gate:
+  `conformance/check.py` against `baseline.json`).
+
+Keep workflows free of JavaScript actions (`actions/checkout`,
+`actions/cache`, …): `rust:latest` has no `node` and not every runner reaches
+GitHub. Check out with plain `git`. Runners are `gitea-runner` 3.5.0 from
+`docker.gitea.com/act_runner` (`gitea/act_runner:latest` on Docker Hub is
+frozen at 0.6.1).
+
+### Conformance
 ```bash
-cargo test --workspace
+CLAWHDF5_PYTHON=.venv/bin/python bash conformance/run.sh --no-fetch   # writes CONFORMANCE.md
 ```
+Reads 697 files of eight pinned corpora with clawhdf5 and h5py and compares
+them object by object (602 ok in the run of 2026-09-28). `CONFORMANCE.md` is
+generated — never hand-edit it (its wording lives in `conformance/report.py`).
+Use `--update-baseline` only after an intended change in results.
+`CONFORMANCE_CACHE` points at an existing corpus cache (`conformance/.cache`,
+about 450 MB). See `conformance/README.md`.
+
+### HDF5 tools (`h5rs`)
+```bash
+cargo run -p clawhdf5-tools -- ls -r file.h5    # also dump [--json], stat, diff, check
+bash scripts/h5rs-fuzz.sh                       # every subcommand over the CVE corpus: no panic/crash/hang
+bash scripts/h5rs-check-ok-files.sh --data      # check passes every fully-read conformance file
+```
+Interop tests compare against h5ls/h5stat/h5dump/h5diff (Debian `hdf5-tools`);
+`dump` must stay byte-identical to h5dump on the test files.
+
+### Remote and browser tests
+- `clawhdf5-remote` tests run a std-only HTTP server
+  (`tests/common/server.rs`, also the `range_server` example);
+  `CLAWHDF5_REMOTE_CORPUS=conformance/.cache/corpus` compares every corpus
+  file over HTTP with `File::open`.
+- wasm: `bash examples/wasm-viewer/test/run.sh` builds the package (needs the
+  `wasm-bindgen` CLI at the crate's exact version) and tests it under Node and
+  headless Chromium (Playwright's download in `~/.cache/ms-playwright` on
+  tank) against `test/serve.py` (range server with request counts). CI has
+  neither, so it runs the native `h5py_interop` and `lazy` tests
+  (`CLAWHDF5_WASM_CORPUS=conformance/.cache/corpus` for the corpus).
+
+### Python bindings
+```bash
+cd crates/clawhdf5-py && maturin develop
+python -m pytest crates/clawhdf5-py/tests      # compares with h5py; editing tests want CLAWHDF5_H5RS=<path to h5rs>
+```
+
+### Benchmarks
+- Search path: `cargo run --release -p clawhdf5-bench --bin search_harness`
+  (`--full`, `--options-study`, `--footprint`, …); reads: `read_harness`,
+  `concurrent_read`; criterion benches with `cargo bench -p <crate>`.
+- Run on an idle machine (1-minute load average below 2; wait otherwise),
+  alternate base and candidate binaries for A/B comparisons, and record date,
+  machine, commit and command with every number in `BENCHMARKS.md`.
+- `BENCHMARKS.md` is written by hand from dated runs; no script regenerates
+  it (the old `scripts/run-benchmarks.sh`, which benchmarked the pre-rename
+  `rustyhdf5-format` and overwrote the file, was removed on 2026-09-28).
 
 ### CLI
 ```bash
 cargo run -p clawhdf5-cli -- --help
-# create, save, search, recall, stats, flush-wal, agents-md, export, snapshot subcommands
-```
-
-### Python bindings
-```bash
-cd crates/clawhdf5-py
-maturin develop
-python -c "import clawhdf5; print(clawhdf5.__version__)"
+# create, save, search, recall, stats, flush-wal, agents-md, export, snapshot, keygen, verify
 ```
 
 ## Integration
-ZeroClaw imports this as a Cargo feature (`clawhdf5` feature flag) to persist agent memory with HNSW vector search for context retrieval.
+- **ClawBrainHub** (`clawverse/clawbrainhub` on git.redclaw.dev) is the one
+  verified consumer: `cbh-core` reads and writes `.brain` files through the
+  facade (`File`, `FileBuilder`, `AttrValue`, `Selection`), `cbh-scanner`
+  uses the facade, and `cbh-cli` uses `clawhdf5_agent::bm25::BM25Index`. It
+  depends on this repo by path (`../clawhdf5`), so changes to those APIs
+  reach it directly. Verified 2026-09-25 against main: builds, and its 204
+  tests pass.
+- OpenClaw and ZeroClaw integrate nothing (see *Standing rules*).

CONFORMANCE.md

@@ -0,0 +1,305 @@
+# clawhdf5 conformance report
+
+Every HDF5 file of eight public corpora (pinned by commit) is read twice — by
+clawhdf5 (`conformance/probe`, the same `clawhdf5-format` calls the facade
+makes) and by h5py/libhdf5 (`conformance/ref.py`) — and the two readings are
+compared object by object: the set of hard-linked objects, each dataset's and
+attribute's shape, and a SHA-256 of its values in a canonical encoding. The
+CVE corpus is also run through `h5dump`. Each side runs under a timeout and an
+address-space limit, so a hang, crash or runaway allocation is recorded, not
+fatal. This file is generated by `conformance/run.sh`; do not edit it by hand.
+
+## Run
+
+| | |
+|---|---|
+| date | 2026-09-28 04:29 UTC |
+| clawhdf5 commit | `bf5a163dcf7fe28d651ada6545d8136ffeffc825` |
+| machine | `tank`: AMD Ryzen 7 7800X3D 8-Core Processor, 16 CPUs, 61 GiB, Linux 7.0.0-34-generic x86_64 |
+| command | `conformance/run.sh --no-fetch --update-baseline` |
+| rustc | rustc 1.98.1 (48a229cea 2026-09-01) |
+| reference | h5py 3.16.0, HDF5 2.0.0, numpy 2.5.3, hdf5plugin 7.1.0, Python 3.14.4 |
+| h5dump | Version 1.14.6 (CVE corpus only) |
+| limits | 20 s timeout (SIGKILL), 4096 MiB address space, per process; 16 files in parallel |
+| runtime | 20 s probing + comparing (5 s fetch/build before it) |
+
+## Results
+
+A file's class is the first that applies:
+
+- **panic / hang / crash / oom** — clawhdf5 panicked (caught per object or not), hit the timeout, died on a signal, or failed an allocation. The CI gate fails on any of these.
+- **h5py-cannot-read** — libhdf5 could not open the file (or itself crashed or hung). Nothing to compare against; most are the deliberately malformed CVE reproducers.
+- **ref-bug** — every difference is an object clawhdf5 refuses that h5py reads only through a libhdf5 bug: the values h5py returns for it change with the reading process's heap, re-checked in every run (see *Reference bugs*).
+- **our-error** — clawhdf5 returned an error for something h5py reads.
+- **mismatch** — both read it, but the shapes, values, object set or attribute set differ.
+- **ok** — every object h5py reads, clawhdf5 reads identically.
+
+| corpus | files | ok | our-error | mismatch | h5py-cannot-read | ref-bug | panic | hang | crash | oom |
+|---|---|---|---|---|---|---|---|---|---|---|
+| NCAS-CMS_pyfive | 33 | 33 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
+| cve_hdf5 | 147 | 113 | 0 | 0 | 32 | 2 | 0 | 0 | 0 | 0 |
+| h5py_data | 4 | 4 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
+| hdf5 | 466 | 405 | 1 | 0 | 60 | 0 | 0 | 0 | 0 | 0 |
+| netcdf-c | 20 | 20 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
+| netcdf4-python | 18 | 18 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
+| usnistgov_h5wasm | 5 | 5 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
+| xarray-data | 4 | 4 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
+| **all** | **697** | **602** | **1** | **0** | **92** | **2** | **0** | **0** | **0** | **0** |
+
+**Our errors and mismatches: 1.** Files not ok: 1 our-error, 92 h5py-cannot-read, 2 ref-bug. 2 object(s) were compared against h5py's values corrected for a known h5py bug (2 identical to clawhdf5's; see *Reference bugs*).
+
+Corpora (fetched by `conformance/fetch-corpus.sh` into the gitignored `conformance/.cache/`):
+
+| corpus | source | commit |
+|---|---|---|
+| hdf5 | https://github.com/HDFGroup/hdf5 | `a3cf1ea82cc7` |
+| cve_hdf5 | https://github.com/HDFGroup/cve_hdf5 | `3fd1f5ae3869` |
+| netcdf-c | https://github.com/Unidata/netcdf-c | `beb7b9585273` |
+| NCAS-CMS_pyfive | https://github.com/NCAS-CMS/pyfive | `8cf07b874913` |
+| usnistgov_h5wasm | https://github.com/usnistgov/h5wasm | `02f6336527d2` |
+| netcdf4-python | https://github.com/Unidata/netcdf4-python | `6e67576d39ae` |
+| xarray-data | https://github.com/pydata/xarray-data | `a35297e9da2c` |
+| h5py_data | https://github.com/h5py/h5py (`h5py/tests/data_files`) | `b2f0347c4200` |
+
+## Panics, hangs, crashes, out-of-memory
+
+None.
+
+## Our-error root causes
+
+Grouped by normalised error message. *files* counts files whose class this cause affects.
+
+| files | objects | error | examples |
+|---:|---:|---|---|
+| 1 | 1 | `ChunkedReadError("…")` | `hdf5/test/testfiles/bad_nbit_parms_walk.h5` |
+
+## Mismatch root causes
+
+None.
+
+## CVE corpus: clawhdf5 vs h5dump vs h5py
+
+The 147 files of [HDFGroup/cve_hdf5](https://github.com/HDFGroup/cve_hdf5) — reproducers for
+published libhdf5 CVEs and fuzzer finds. *read* = produced output (possibly with per-object
+errors), *error* = refused cleanly. h5dump exits non-zero on any error anywhere in a file, so
+its read/error split is not comparable with the other two rows; the panic, crash, hang and oom
+columns are.
+
+| tool | read | error | panic | crash | hang | oom |
+|---|---:|---:|---:|---:|---:|---:|
+| clawhdf5 | 121 | 26 | 0 | 0 | 0 | 0 |
+| h5dump 1.14.6 | 16 | 129 | 0 | 2 | 0 | 0 |
+| h5py 3.16.0 / HDF5 2.0.0 | 115 | 31 | 0 | 1 | 0 | 0 |
+
+<details><summary>Per-file outcomes</summary>
+
+| file | h5dump | h5py | clawhdf5 | class |
+|---|---|---|---|---|
+| cvefiles/cve-2016-4330.h5 | error exit | read 3 obj, 1 errors | read 3 obj, 1 errors | ok |
+| cvefiles/cve-2016-4331.h5 | error exit | read 25 obj, 1 errors | read 25 obj, 1 errors | ok |
+| cvefiles/cve-2016-4332-mtime-new.h5 | error exit | read 25 obj, 1 errors | read 25 obj, 1 errors | ok |
+| cvefiles/cve-2016-4332-mtime.h5 | error exit | read 4 obj, 3 errors | read 4 obj, 3 errors | ok |
+| cvefiles/cve-2016-4332-stab.h5 | error exit | open error | read 1 obj, 1 errors | h5py-cannot-read |
+| cvefiles/cve-2016-4333.h5 | error exit | read 3 obj, 1 errors | read 3 obj, 1 errors | ok |
+| cvefiles/cve-2017-17505.h5 | error exit | read 2 obj, 1 errors | read 2 obj, 1 errors | ok |
+| cvefiles/cve-2017-17506.h5 | error exit | read 2 obj, 1 errors | read 2 obj, 1 errors | ok |
+| cvefiles/cve-2017-17507.h5 | error exit | read 2 obj, 1 errors | read 2 obj, 1 errors | ok |
+| cvefiles/cve-2017-17508.h5 | error exit | read 2 obj, 1 errors | read 2 obj, 1 errors | ok |
+| cvefiles/cve-2017-17509.h5 | error exit | read 2 obj, 1 errors | read 2 obj, 1 errors | ok |
+| cvefiles/cve-2018-11202.h5 | error exit | read 2 obj, 1 errors | read 2 obj, 1 errors | ok |
+| cvefiles/cve-2018-11203.h5 | error exit | read 2 obj, 1 errors | read 2 obj, 1 errors | ok |
+| cvefiles/cve-2018-11204.h5 | error exit | read 2 obj, 1 errors | read 2 obj, 1 errors | ok |
+| cvefiles/cve-2018-11205.h5 | error exit | read 2 obj, 1 errors | read 2 obj, 1 errors | ok |
+| cvefiles/cve-2018-11206-new.h5 | error exit | read 3 obj, 1 errors | read 3 obj, 1 errors | ok |
+| cvefiles/cve-2018-11206-old.h5 | error exit | read 3 obj, 1 errors | read 3 obj, 1 errors | ok |
+| cvefiles/cve-2018-11207.h5 | error exit | read 2 obj, 1 errors | read 2 obj, 1 errors | ok |
+| cvefiles/cve-2018-13866.h5 | error exit | open error | open error | h5py-cannot-read |
+| cvefiles/cve-2018-13867.h5 | error exit | read 1 obj, 1 errors | read 1 obj, 1 errors | ok |
+| cvefiles/cve-2018-13868.h5 | error exit | read 3 obj, 1 errors | read 3 obj, 1 errors | ok |
+| cvefiles/cve-2018-13869.h5 | error exit | read 1 obj, 1 errors | read 1 obj, 1 errors | ok |
+| cvefiles/cve-2018-13870.h5 | error exit | read 1 obj, 1 errors | read 1 obj, 1 errors | ok |
+| cvefiles/cve-2018-13871.h5 | error exit | read 2 obj | read 2 obj | ok |
+| cvefiles/cve-2018-13872.h5 | error exit | read 1 obj, 1 errors | read 1 obj, 1 errors | ok |
+| cvefiles/cve-2018-13873.h5 | error exit | read 1 obj, 1 errors | read 1 obj, 1 errors | ok |
+| cvefiles/cve-2018-13874.h5 | error exit | open error | open error | h5py-cannot-read |
+| cvefiles/cve-2018-13875.h5 | error exit | read 3 obj, 1 errors | read 3 obj, 1 errors | ok |
+| cvefiles/cve-2018-13876.h5 | error exit | open error | open error | h5py-cannot-read |
+| cvefiles/cve-2018-14031.h5 | error exit | read 3 obj, 1 errors | read 3 obj, 1 errors | ok |
+| cvefiles/cve-2018-14033.h5 | error exit | read 3 obj, 1 errors | read 3 obj, 1 errors | ok |
+| cvefiles/cve-2018-14034.h5 | error exit | read 1 obj, 2 errors | read 1 obj | ok |
+| cvefiles/cve-2018-14035.h5 | error exit | read 3 obj, 1 errors | read 3 obj, 1 errors | ok |
+| cvefiles/cve-2018-14460.h5 | error exit | read 3 obj, 2 errors | read 3 obj, 2 errors | ok |
+| cvefiles/cve-2018-15671.h5 | ok | read 1 obj | read 1 obj | ok |
+| cvefiles/cve-2018-15672.h5 | error exit | read 2 obj, 1 errors | read 2 obj, 1 errors | ok |
+| cvefiles/cve-2018-16438.h5 | error exit | read 1 obj, 1 errors | read 1 obj | ok |
+| cvefiles/cve-2018-17233.h5 | error exit | read 6 obj, 1 errors | read 6 obj, 1 errors | ok |
+| cvefiles/cve-2018-17234.h5 | error exit | read 6 obj, 1 errors | read 6 obj, 1 errors | ok |
+| cvefiles/cve-2018-17237.h5 | error exit | read 2 obj, 1 errors | read 2 obj, 1 errors | ok |
+| cvefiles/cve-2018-17432.h5 | error exit | read 3 obj, 1 errors | read 3 obj, 1 errors | ok |
+| cvefiles/cve-2018-17433 | error exit | open error | open error | h5py-cannot-read |
+| cvefiles/cve-2018-17434.h5 | error exit | read 3 obj, 1 errors | read 3 obj, 1 errors | ok |
+| cvefiles/cve-2018-17435.h5 | error exit | read 3 obj, 1 errors | read 3 obj, 1 errors | ok |
+| cvefiles/cve-2018-17436 | error exit | open error | open error | h5py-cannot-read |
+| cvefiles/cve-2018-17437.h5 | error exit | read 3 obj, 1 errors | read 3 obj, 1 errors | ok |
+| cvefiles/cve-2018-17438 | error exit | read 3 obj, 1 errors | read 3 obj, 1 errors | ok |
+| cvefiles/cve-2018-17439 | error exit | read 3 obj, 1 errors | read 3 obj, 1 errors | ok |
+| cvefiles/cve-2019-8396.h5 | error exit | read 3 obj, 2 errors | read 3 obj, 2 errors | ok |
+| cvefiles/cve-2019-8397.h5 | error exit | read 2 obj, 1 errors | read 2 obj, 1 errors | ok |
+| cvefiles/cve-2019-8398.h5 | error exit | read 2 obj, 1 errors | read 2 obj, 1 errors | ok |
+| cvefiles/cve-2019-9151.h5 | error exit | read 3 obj, 1 errors | read 3 obj, 2 errors | ok |
+| cvefiles/cve-2019-9152.h5 | error exit | read 3 obj, 1 errors | read 3 obj, 1 errors | ok |
+| cvefiles/cve-2020-10809 | error exit | open error | open error | h5py-cannot-read |
+| cvefiles/cve-2020-10810.h5 | error exit | open error | open error | h5py-cannot-read |
+| cvefiles/cve-2020-10811.h5 | error exit | read 25 obj, 1 errors | read 25 obj, 1 errors | ok |
+| cvefiles/cve-2020-10812.h5 | error exit | open error | open error | h5py-cannot-read |
+| cvefiles/cve-2020-18232.h5 | error exit | read 3 obj, 2 errors | read 3 obj, 2 errors | ok |
+| cvefiles/cve-2020-18494.h5 | ok | read 2 obj | read 2 obj | ok |
+| cvefiles/cve-2021-36977.h5 | error exit | read 1 obj, 1 errors | read 1 obj, 1 errors | ok |
+| cvefiles/cve-2021-37501.h5 | error exit | read 18 obj, 1 errors | read 18 obj, 1 errors | ok |
+| cvefiles/cve-2021-45829.h5 | error exit | read 1 obj, 2 errors | read 1 obj, 1 errors | ok |
+| cvefiles/cve-2021-45830.h5 | error exit | open error | open error | h5py-cannot-read |
+| cvefiles/cve-2021-45833.h5 | error exit | read 2 obj, 1 errors | read 2 obj, 1 errors | ok |
+| cvefiles/cve-2021-46242.h5 | error exit | open error | read 1 obj, 1 errors | h5py-cannot-read |
+| cvefiles/cve-2021-46243.h5 | error exit | read 2 obj, 1 errors | read 2 obj, 1 errors | ok |
+| cvefiles/cve-2021-46244.h5 | error exit | read 2 obj, 1 errors | read 2 obj, 1 errors | ok |
+| cvefiles/cve-2024-29157.h5 | error exit | read 4 obj, 7 errors | read 4 obj, 7 errors | ok |
+| cvefiles/cve-2024-29158.h5 | ok | read 3 obj, 1 errors | read 3 obj, 1 errors | ok |
+| cvefiles/cve-2024-29159.h5 | error exit | read 2 obj, 1 errors | read 2 obj, 1 errors | ok |
+| cvefiles/cve-2024-29160.h5 | error exit | read 4 obj, 1 errors | read 4 obj, 1 errors | ok |
+| cvefiles/cve-2024-29161.h5 | error exit | read 1 obj, 1 errors | read 1 obj, 1 errors | ok |
+| cvefiles/cve-2024-29162.h5 | error exit | read 17 obj, 4 errors | read 17 obj, 4 errors | ok |
+| cvefiles/cve-2024-29163.h5 | error exit | read 7 obj, 1 errors | read 7 obj, 1 errors | ok |
+| cvefiles/cve-2024-29164.h5 | ok | read 3 obj | read 3 obj | ok |
+| cvefiles/cve-2024-29165.h5 | error exit | read 2 obj, 1 errors | read 2 obj, 1 errors | ok |
+| cvefiles/cve-2024-29166.h5 | error exit | read 17 obj, 2 errors | read 17 obj | ok |
+| cvefiles/cve-2024-32605.h5 | ok | read 6 obj, 1 errors | read 6 obj, 1 errors | ok |
+| cvefiles/cve-2024-32606.h5 | error exit | read 2 obj, 1 errors | read 2 obj, 1 errors | ok |
+| cvefiles/cve-2024-32607-1.h5 | ok | read 10 obj | read 10 obj | ok |
+| cvefiles/cve-2024-32607-2.h5 | error exit | read 9 obj, 1 errors | read 9 obj, 1 errors | ok |
+| cvefiles/cve-2024-32608.h5 | error exit | read 6 obj, 1 errors | read 6 obj, 1 errors | ok |
+| cvefiles/cve-2024-32609.h5 | error exit | SIGSEGV | read 3 obj, 1 errors | h5py-cannot-read |
+| cvefiles/cve-2024-32610.h5 | error exit | read 2 obj, 1 errors | read 2 obj, 1 errors | ok |
+| cvefiles/cve-2024-32611.h5 | ok | read 6 obj | read 6 obj | ok |
+| cvefiles/cve-2024-32612.h5 | ok | read 3 obj | read 3 obj | ok |
+| cvefiles/cve-2024-32613.h5 | error exit | read 7 obj, 1 errors | read 7 obj, 1 errors | ok |
+| cvefiles/cve-2024-32614.h5 | error exit | read 25 obj, 2 errors | read 25 obj, 2 errors | ok |
+| cvefiles/cve-2024-32615.h5 | error exit | read 4 obj, 1 errors | read 4 obj, 1 errors | ok |
+| cvefiles/cve-2024-32616.h5 | error exit | read 10 obj, 7 errors | read 10 obj, 6 errors | ok |
+| cvefiles/cve-2024-32617.h5 | error exit | read 1 obj, 1 errors | read 1 obj, 1 errors | ok |
+| cvefiles/cve-2024-32618.h5 | error exit | read 3 obj, 1 errors | read 3 obj, 1 errors | ok |
+| cvefiles/cve-2024-32619.h5 | error exit | read 3 obj, 2 errors | read 3 obj, 2 errors | ok |
+| cvefiles/cve-2024-32620.h5 | error exit | read 1 obj, 1 errors | read 1 obj, 1 errors | ok |
+| cvefiles/cve-2024-32621.h5 | ok | read 3 obj, 1 errors | read 3 obj, 1 errors | ok |
+| cvefiles/cve-2024-32622.h5 | ok | read 3 obj, 1 errors | read 3 obj, 1 errors | ok |
+| cvefiles/cve-2024-32623.h5 | ok | read 6 obj | read 6 obj | ok |
+| cvefiles/cve-2024-32624.h5 | error exit | read 6 obj, 1 errors | read 6 obj, 1 errors | ok |
+| cvefiles/cve-2024-33873.h5 | error exit | read 4 obj, 1 errors | read 4 obj, 1 errors | ok |
+| cvefiles/cve-2024-33874.h5 | ok | read 6 obj, 1 errors | read 6 obj, 1 errors | ok |
+| cvefiles/cve-2024-33875.h5 | ok | read 2 obj | read 2 obj | ok |
+| cvefiles/cve-2024-33876.h5 | ok | read 3 obj, 1 errors | read 3 obj, 1 errors | ok |
+| cvefiles/cve-2024-33877.h5 | error exit | read 8 obj, 1 errors | read 8 obj, 1 errors | ok |
+| cvefiles/cve-2025-2153.h5 | error exit | open error | read 1 obj, 1 errors | h5py-cannot-read |
+| cvefiles/cve-2025-2308.h5 | error exit | read 25 obj, 1 errors | read 25 obj, 2 errors | ref-bug |
+| cvefiles/cve-2025-2309.h5 | ok | read 6 obj, 1 errors | read 6 obj | ok |
+| cvefiles/cve-2025-2310.h5 | error exit | read 24 obj, 8 errors | read 24 obj, 8 errors | ok |
+| cvefiles/cve-2025-2912.h5 | error exit | open error | open error | h5py-cannot-read |
+| cvefiles/cve-2025-2913.h5 | error exit | open error | open error | h5py-cannot-read |
+| cvefiles/cve-2025-2914.h5 | error exit | open error | open error | h5py-cannot-read |
+| cvefiles/cve-2025-2915.h5 | error exit | open error | open error | h5py-cannot-read |
+| cvefiles/cve-2025-2923.h5 | error exit | open error | open error | h5py-cannot-read |
+| cvefiles/cve-2025-2924.h5 | error exit | read 1 obj, 1 errors | read 1 obj, 1 errors | ok |
+| cvefiles/cve-2025-2925.h5 | error exit | read 1 obj, 1 errors | read 1 obj, 1 errors | ok |
+| cvefiles/cve-2025-2926.h5 | error exit | open error | open error | h5py-cannot-read |
+| cvefiles/cve-2025-44904.h5 | error exit | read 25 obj, 1 errors | read 25 obj, 2 errors | ref-bug |
+| cvefiles/cve-2025-44905.h5 | error exit | read 25 obj, 3 errors | read 25 obj, 3 errors | ok |
+| cvefiles/cve-2025-6269-1.h5 | error exit | read 1 obj, 1 errors | read 1 obj, 1 errors | ok |
+| cvefiles/cve-2025-6269-2.h5 | error exit | read 1 obj, 1 errors | read 1 obj, 1 errors | ok |
+| cvefiles/cve-2025-6269-3.h5 | error exit | read 1 obj, 1 errors | read 1 obj, 1 errors | ok |
+| cvefiles/cve-2025-6269-4.h5 | error exit | read 1 obj, 1 errors | read 1 obj, 1 errors | ok |
+| cvefiles/cve-2025-6270-1.h5 | error exit | open error | open error | h5py-cannot-read |
+| cvefiles/cve-2025-6270-2.h5 | error exit | open error | open error | h5py-cannot-read |
+| cvefiles/cve-2025-6270-3.h5 | error exit | open error | open error | h5py-cannot-read |
+| cvefiles/cve-2025-6516.h5 | error exit | read 1 obj, 1 errors | read 1 obj, 1 errors | ok |
+| cvefiles/cve-2025-6750.h5 | error exit | open error | read 1 obj, 1 errors | h5py-cannot-read |
+| cvefiles/cve-2025-6816.h5 | error exit | open error | open error | h5py-cannot-read |
+| cvefiles/cve-2025-6817.h5 | error exit | open error | open error | h5py-cannot-read |
+| cvefiles/cve-2025-6818.h5 | error exit | open error | open error | h5py-cannot-read |
+| cvefiles/cve-2025-6856.h5 | error exit | open error | open error | h5py-cannot-read |
+| cvefiles/cve-2025-6857.h5 | error exit | read 1 obj, 1 errors | read 1 obj, 1 errors | ok |
+| cvefiles/cve-2025-6858.h5 | SIGSEGV | open error | open error | h5py-cannot-read |
+| cvefiles/cve-2025-7067.h5 | error exit | read 1 obj, 1 errors | read 1 obj, 1 errors | ok |
+| cvefiles/cve-2025-7068.h5 | error exit | open error | open error | h5py-cannot-read |
+| cvefiles/cve-2025-7069.h5 | error exit | open error | open error | h5py-cannot-read |
+| cvefiles/cve-2026-26200.h5 | error exit | read 1 obj, 1 errors | read 1 obj, 1 errors | ok |
+| cvefiles/cve-2026-34734.h5 | error exit | read 2 obj, 1 errors | read 2 obj, 1 errors | ok |
+| cvefiles/cve-2026-92627.h5 | error exit | read 2 obj, 1 errors | read 2 obj, 1 errors | ok |
+| cvefiles/unknown-1.h5 | error exit | read 11 obj, 1 errors | read 11 obj, 1 errors | ok |
+| fuzzerfiles/gh-4431-poc-03.h5 | error exit | read 1 obj | read 1 obj | ok |
+| fuzzerfiles/gh-4432-poc-05.h5 | SIGSEGV | read 1 obj, 1 errors | read 1 obj, 1 errors | ok |
+| fuzzerfiles/gh-4433-poc-08.h5 | error exit | read 1 obj, 1 errors | read 1 obj, 1 errors | ok |
+| fuzzerfiles/gh-4434-poc-09.h5 | error exit | open error | read 1 obj, 1 errors | h5py-cannot-read |
+| fuzzerfiles/gh-4435-poc-10.h5 | error exit | read 1 obj, 1 errors | read 1 obj, 1 errors | ok |
+| fuzzerfiles/gh-4585.h5 | error exit | open error | open error | h5py-cannot-read |
+| fuzzerfiles/gh_2649_flawed.h5 | error exit | read 9 obj, 1 errors | read 9 obj, 1 errors | ok |
+| fuzzerfiles/gh_2649_plain_model.h5 | ok | read 10 obj | read 10 obj | ok |
+
+</details>
+
+## Reference bugs
+
+### Objects h5py reads only through a libhdf5 bug (*ref-bug*)
+
+clawhdf5 refuses these objects; h5py 3.16 / HDF5 2.0 returns values for them. `conformance/ref_bugs.py`
+re-reads each with h5py in six fresh processes whose heaps differ (h5py imported before numpy, three
+times and twice more with `MALLOC_PERTURB_`, and numpy imported first). Values the file determines
+come out the same every time; these do not, so they are memory libhdf5 over-reads, not the file's
+data. A file is *ref-bug* only while every one of its differences is such an object confirmed in
+the same run; an object that reads the same every time goes back to *our-error*. Reproducer:
+`python conformance/ref_bugs.py conformance/.cache/corpus` (prints every read's outcome).
+
+| file | object | distinct results in 6 reads | confirmed | what goes wrong |
+|---|---|---:|---|---|
+| `cve_hdf5/cvefiles/cve-2025-2308.h5` | `/Scale_offset_long_long_data_le` | 6 | yes | the first chunk records minbits 11: its 12 values need 17 bytes of codes, and the 26-byte chunk holds 5 after its 21-byte header; libhdf5's scale-offset decoder reads past its buffer, and develop refuses the chunk ("Buffer too short") |
+| `cve_hdf5/cvefiles/cve-2025-44904.h5` | `/Scale_offset_float_data_le` | 6 | yes | unfiltered chunks stored as 38 and 37 bytes for 48-byte chunks: 1.14/2.0 read the stored bytes into a buffer of that size and use it as the whole chunk (H5D__chunk_lock), so the rest is heap memory; develop refuses them ("incorrect chunk size returned from index for unfiltered chunk") |
+| `hdf5/test/testfiles/bad_nbit_parms_walk.h5` | `/Nbit_int_data_le` | 1 | **no** | the N-Bit parameter list holds 7 values (cd_values[0] = 7) where an integer needs 8: the decoder takes the bit offset from cd_values[7], past the list; libhdf5's own test (`test_filter_bad_params`, test/dsets.c on develop) requires the read to fail |
+
+### Values corrected for a known h5py bug
+
+- **h5py big-endian variable-length sequences.** h5py returns the elements of a VL sequence
+  whose base type is big-endian with the file's big-endian bytes but a native (little-endian)
+  numpy dtype: a `h5py.vlen_dtype(np.dtype('>f4'))` dataset holding `[1.0, 2.0]` reads back as
+  `[4.6e-41, 9.0e-44]`; `h5dump` prints the file's values. `ref.py` checks that the installed
+  h5py still does this (by writing and reading exactly that dataset in memory) and, if so,
+  relabels such elements with the file's byte order before hashing, so the values are still
+  compared. Corrected objects: `NCAS-CMS_pyfive/tests/data/attr_datatypes.hdf5` `/@vlen_uint64` (same as clawhdf5), `hdf5/tools/test/testfiles/tcomplex_be.h5` `/VariableLengthDatasetFloatComplex` (same as clawhdf5).
+
+## Other comparison rules
+
+- **Non-IEEE floats and partial-precision integers (N-Bit).** libhdf5 converts a float whose
+  bit layout is not IEEE (e.g. `H5Tset_precision` for the N-Bit filter) or an integer with a
+  bit offset / reduced precision into the plain numpy type of the same size. The probe
+  compares such values as converted numbers, not raw file bytes (before 2026-09-25 it compared
+  raw bytes, which reported every N-Bit float dataset as a mismatch).
+- **Types h5py widens.** Where h5py reads a type into a numpy type of a different size
+  (FP8 -> float16, bfloat16 -> float32, x87 long double -> float128) the values are not
+  compared (shape and presence still are): dataset file type size 1 -> numpy float16 (2) (15x), attr file type size 1 -> numpy float16 (2) (15x), dataset file type size 2 -> numpy float32 (4) (2x), dataset file type size 8 -> numpy float128 (16) (1x), dataset file type size 12 -> numpy float128 (16) (1x), attr file type size 2 -> numpy float32 (4) (1x), dataset file type size 2 -> numpy >f4 (4) (1x), attr file type size 2 -> numpy >f4 (4) (1x).
+- **References** are compared by presence only (`R`), not by target.
+
+## Objects h5py fails on but clawhdf5 reads
+
+- 19 x `OSError: Can't synchronously read data (no appropriate function for conversion path)`
+- 1 x `TypeError: unhandled dtype kind M (dtype('…'))`
+- 1 x `TypeError: No NumPy equivalent for TypeTimeID exists`
+- 1 x `ValueError: Insufficient precision in available types to represent (N, N, N, N, N)`
+
+## Reproduce
+
+```sh
+# needs: Rust, python3 with h5py numpy hdf5plugin (conformance/requirements.txt), h5dump (hdf5-tools), git
+CLAWHDF5_PYTHON=/path/to/venv/bin/python conformance/run.sh
+```
+
+The corpus (about 450 MB of sparse checkouts) is cached in `conformance/.cache/`; results for
+every file, both sides' raw JSON and stderr, are in `conformance/.cache/results/`.
+`conformance/baseline.json` holds the ok files the nightly CI job (`.gitea/workflows/conformance.yml`)
+must keep; `conformance/run.sh --update-baseline` rewrites it.

Cargo.toml

@@ -16,13 +16,19 @@ members = [
     "crates/clawhdf5-cli",
     "crates/clawhdf5-napi",
     "crates/clawhdf5-bench",
+    "crates/clawhdf5-tools",
+    "crates/clawhdf5-wasm",
+    "crates/clawhdf5-remote",
     "crates/libaec-sys",
 ]
 resolver = "2"
 
 [workspace.package]
-version = "2.5.0"
+version = "2.7.0"
 edition = "2024"
+# Oldest toolchain that builds the whole workspace; CI checks it. wgpu (in
+# clawhdf5-gpu) requires 1.92.
+rust-version = "1.92"
 license = "MIT"
 repository = "https://git.redclaw.dev/quantumclaw/clawhdf5"
 
@@ -31,3 +37,12 @@ tempfile = "3"
 criterion = { version = "0.5", features = ["html_reports"] }
 half = "2.7"
 serde = { version = "1", features = ["derive"] }
+
+# The browser build of clawhdf5-wasm (examples/wasm-viewer/build.sh): size
+# over speed, whole-program optimisation. Native profiles are unaffected.
+[profile.wasm-release]
+inherits = "release"
+opt-level = "s"
+lto = true
+codegen-units = 1
+panic = "abort"

README.md

@@ -1,582 +1,452 @@
-# ClawhDF5
+# clawhdf5
 
-**The memory layer AI agents deserve. One file. Pure Rust. Zero C dependencies.**
+**A pure-Rust HDF5 reader, writer and in-place editor — no libhdf5, and no
+C by default — with an agent-memory store built on it.**
 
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](LICENSE)
-[![Rust](https://img.shields.io/badge/rust-1.75%2B-orange.svg)](https://www.rust-lang.org)
-[![Tests](https://img.shields.io/badge/tests-1650%2B%20passing-brightgreen.svg)](#performance)
-[![LongMemEval](https://img.shields.io/badge/LongMemEval%20oracle-Turn--Level%20Hit@5%2084%25%20BM25--only-blue.svg)](BENCHMARKS.md#longmemeval-results)
-[![Footprint](https://img.shields.io/badge/footprint-6.5%20KB%2Frecord-lightgrey.svg)](BENCHMARKS.md#memory-footprint)
+[![Rust](https://img.shields.io/badge/rust-1.92%2B-orange.svg)](https://www.rust-lang.org)
+[![Conformance](https://img.shields.io/badge/conformance-602%2F697%20files%20identical%20to%20h5py-brightgreen.svg)](CONFORMANCE.md)
+[![LongMemEval](https://img.shields.io/badge/LongMemEval__s-turn%20Hit@5%2081.4%25%20hybrid-blue.svg)](BENCHMARKS.md#longmemeval-results)
 
-ClawHDF5 is a pure-Rust HDF5 implementation combined with a research-grade agent memory engine. It gives AI agents persistent, searchable, cryptographically verifiable memory — all stored in a single portable file.
+clawhdf5 implements the HDF5 file format from the specification, in Rust.
+It reads superblocks v0–3, every group and chunk-index structure libhdf5
+writes, the standard filters and the common plugin filters,
+variable-length data and virtual datasets, and follows files a SWMR writer
+is appending to. It reads files from libhdf5, h5py and netCDF-4 and writes
+files they read, in the HDF5 1.10 format or, on request, in one HDF5 1.8
+reads. The same library opens files over HTTP and in object
+stores by range requests, runs in the browser as WebAssembly, and has
+Python bindings with an h5py-shaped API.
 
-> **Two things live here:**
-> - **A general-purpose, pure-Rust HDF5 library** — zero C dependencies, NetCDF-4 support, SIMD/GPU acceleration. See the **[Crate Map](#crate-map)** and **[BENCHMARKS.md](BENCHMARKS.md)** for the libhdf5 head-to-head numbers.
-> - **An agent memory layer built on top of it** — vector search, knowledge graph, hippocampal-style consolidation, in `clawhdf5-agent`.
+Two things live in this repository:
 
-```
-cargo add clawhdf5                          # core HDF5 read/write, no agent layer
-cargo add clawhdf5-agent --features agent   # + agent memory layer
+- **The HDF5 library** — the `clawhdf5` crate and its parts, `h5rs`
+  command-line tools, Python, WebAssembly and NetCDF-4 layers.
+- **Agent memory** (`clawhdf5-agent`) — a single-file store for AI agents
+  (HNSW + BM25 hybrid search, write-ahead log, signed checkpoints) whose
+  files are ordinary HDF5. See [docs/agent-memory.md](docs/agent-memory.md).
+
+Nothing is published to crates.io or PyPI yet: use it
+[from git or a checkout](#install).
+
+## Contents
+
+- [Evidence](#evidence) — conformance, robustness on hostile files, speed
+- [What is supported](#what-is-supported) — the feature matrix
+- [Install](#install) · [Quick start: Rust](#quick-start-rust) · [Quick start: Python](#quick-start-python)
+- [Remote files, the browser, SWMR](#remote-files-the-browser-swmr) · [h5rs tools](#h5rs-tools)
+- [Agent memory](#agent-memory) · [Crate map](#crate-map) · [Building and testing](#building-and-testing)
+- [Documentation](#documentation) · [Who uses it](#who-uses-it)
+
+## Evidence
+
+**Conformance.** Every file of eight public corpora — the libhdf5 source
+tree's test files, the HDF Group's
+[CVE reproducer corpus](https://github.com/HDFGroup/cve_hdf5), netcdf-c,
+netcdf4-python, pyfive, h5wasm, h5py's and xarray's data files, 697 files
+in all, pinned by commit — is read by clawhdf5 and by h5py/libhdf5 and
+compared object by object (object set, shapes, a SHA-256 of every
+dataset's and attribute's values). Run of 2026-09-28 on tank, h5py 3.16 /
+HDF5 2.0 ([CONFORMANCE.md](CONFORMANCE.md)):
+
+| files | ok (identical to h5py) | mismatch | libhdf5 cannot open | ref-bug¹ | our-error¹ | panic / hang / crash / OOM |
+|---:|---:|---:|---:|---:|---:|---:|
+| 697 | **602** | **0** | 92 | 2 | 1 | **0** |
+
+¹ The three remaining objects are corrupt data (scale-offset codes past
+the end of a chunk, short unfiltered chunks, an N-Bit parameter list one
+value short) that HDF5 2.0 returns only by reading past a buffer;
+clawhdf5 refuses them, as libhdf5's development branch and its own
+`test_filter_bad_params` do. Details and evidence in
+[CONFORMANCE.md § Reference bugs](CONFORMANCE.md#reference-bugs); the
+N-Bit file is counted as ref-bug or our-error depending on the run
+([known-issues](docs/known-issues.md#conformance-bad_nbit_parms_walkh5-flips-between-ref-bug-and-our-error)). The
+run is a nightly CI job (`.gitea/workflows/conformance.yml`) that fails on
+any panic, hang or crash, or on an ok file that stops being ok.
+
+**Robustness on hostile files.** On the 147 CVE and fuzzer files
+([CONFORMANCE.md § CVE corpus](CONFORMANCE.md#cve-corpus-clawhdf5-vs-h5dump-vs-h5py)):
+
+| tool | panic | crash | hang | OOM |
+|---|---:|---:|---:|---:|
+| clawhdf5 | 0 | 0 | 0 | 0 |
+| h5dump 1.14.6 | 0 | 2 | 0 | 0 |
+| h5py 3.16.0 / HDF5 2.0.0 | 0 | 1 | 0 | 0 |
+
+Sizes and addresses read from a file are checked before use
+(overflow-checked arithmetic, fallible allocation on the chunked read
+paths, bounded recursion in B-trees and object-header chains), and
+`scripts/h5rs-fuzz.sh` runs every `h5rs` subcommand over the corpus
+looking for panics, crashes and hangs.
+
+**Reads from many threads.** A `File` is `Send + Sync` and there is no
+library-wide lock, so one open file serves many threads. Full reads of 64
+deflate-compressed 64 MiB datasets, each read decoding on its calling
+thread (`concurrent_read --decode-threads 1`), tank (Ryzen 7 7800X3D,
+16 threads), 2026-09-26, commit `c5334b1`
+([BENCHMARKS.md](BENCHMARKS.md#results-after-in-place-chunk-decoding-2026-09-26-tank-c5334b1)):
+
+| threads | clawhdf5, one `File` | h5py, threads | h5py, processes | clawhdf5 / h5py processes |
+|---:|---:|---:|---:|---:|
+| 1 | 670 MB/s | 410 MB/s | 397 MB/s | 1.69x |
+| 16 | 4944 MB/s | 390 MB/s | 3135 MB/s | 1.58x |
+
+That run was noisier than others on the same machine, so compare ratios
+within it rather than MB/s across runs. A contiguous (uncompressed) full
+read on one thread ran at 6718 MB/s against h5py's 5545 in the same run.
+
+**Against libhdf5 1.14.6 from Rust**, tank, 2026-08-03
+([BENCHMARKS.md § Independent Validation](BENCHMARKS.md#independent-validation-tank-ryzen-7-7800x3d-2026-08-03)):
+sequential read of 100K `f32` 23.3 µs vs 63.6 µs (2.7x); 128 attribute
+writes 85.2 µs vs 877 µs (10.3x); 64 group creates 130 µs vs 1.37 ms
+(10.6x); a 512×512 `f32` chunked deflate-6 write 1.44 ms vs 65.0 ms
+(re-measured 2026-09-23 with the pure-Rust deflate: 1.46 ms vs 51.4 ms,
+35x); a 100K `f32` sequential write is a tie. The writer (`FileBuilder`)
+assembles a file in memory and writes it once, which is part of that
+difference; read the caveats in [BENCHMARKS.md](BENCHMARKS.md#caveats)
+before quoting these.
+
+## What is supported
+
+Limits and open issues, with dates, are in
+[docs/known-issues.md](docs/known-issues.md).
+
+| Area | Supported | Read only | Not supported |
+|---|---|---|---|
+| **File format** | Superblock v0–v3, user blocks, v1/v2 object headers; writing the HDF5 1.10 format (default) or, with `libver_bounds(LibVer::V18, LibVer::V18)`, files HDF5 1.8 reads (checked with HDF5 1.8.23) | Metadata cache images | Writing the pre-1.8 format (version-0 superblock, symbol-table groups) |
+| **Groups and links** | Symbol-table, compact and dense groups (tested to 100 000 links), creation order, soft and hard links; writing external links | | Following external links (explicit error); user-defined links are skipped |
+| **Datatypes** | Integers and IEEE floats of every width and byte order (incl. `f16`), enums, compounds (every version, incl. HDF5 2.0's v5), arrays, fixed-length strings, opaque, complex: h5py's `{r, i}` compound (`with_complex_f64_data`) and HDF5 2.0's native class 11 (`with_native_complex_f64_data`, opt-in: only libhdf5 2.0+ reads it; reads surface it as `{r, i}`) | Variable-length strings and sequences, object references; HDF5 2.x's small floats (bfloat16, FP8 E4M3/E5M2, FP6 E2M3/E3M2, FP4 E2M1: every bit pattern decoded as libhdf5 2.2.0 decodes it) and other non-IEEE floats up to 64 bits | Writing variable-length data; writing non-IEEE floats; decoding region and attribute references; x87 long double and binary128 |
+| **Layouts and chunk indexes** | Compact, contiguous and chunked; chunk indexes single chunk, Fixed Array, Extensible Array and v2 B-tree (the writer picks one as libhdf5 does), and v1 B-tree (every chunked dataset under the 1.8 bound, split as libhdf5 splits it); chunks of 4 GiB or more (HDF5 2.0's layout message version 5); fill values; resizable datasets; virtual datasets (read limits in known-issues) | The implicit chunk index (the editor also changes it) | External raw data files (explicit error); chunk dimensions of 2^32 or more |
+| **Filters** | deflate (pure-Rust zlib-rs), shuffle, Fletcher-32, LZ4 (opt-in), Zstd (C, opt-in); plugins LZF, bitshuffle, bzip2, Blosc 1 | N-Bit, scale-offset, SZIP (C, opt-in); plugins Blosc2 and ZFP | Other filter IDs, unless you register a codec (`filter_registry::register_filter`) |
+| **Editing in place** | `FileEditor`: overwrite values, grow and shrink chunked datasets (every index), set attributes (compact and dense), in files from h5py or clawhdf5 | | Creating or deleting objects in an existing file; deleting attributes; new chunks in implicit indexes; VL data; rewriting chunks of 4 GiB or more; filters this build cannot encode (refused before any write) |
+| **Access** | Local files (mmap or buffered), bytes in memory, any `Storage` backend, HTTP(S) and S3/GCS/Azure via `clawhdf5-remote`, SWMR reading (`File::open_swmr`, `Dataset::refresh`) | Remote files and the browser are read-only | SWMR writing; remote SWMR; MPI collective I/O (`clawhdf5-io`'s `mpi-io` reads on one rank and broadcasts) |
+| **Bindings** | Python (read, `'w'` for numeric and complex arrays, `'r+'` editing, URLs), NetCDF-4 (CF scale/offset/fill) | WebAssembly (`open(bytes)`, `openUrl`); no Zstd/SZIP/pcodec, no compound, reference, opaque, bitfield, time or VL-sequence datasets | Node.js (the package does not work; see known-issues) |
+
+Plugin filters other than LZF are cargo features (`bitshuffle`, `bzip2`,
+`blosc`, `blosc2`, `zfp`, or `plugin-filters` for all of them), all pure
+Rust; h5py + hdf5plugin read what clawhdf5 writes with them, and ZFP decodes
+bit-exact against hdf5plugin 7.1. `pcodec` (opt-in) uses a private filter ID
+that only clawhdf5 reads.
+
+**C dependencies, precisely.** The core crates build no C by default: no
+libhdf5, and deflate is [zlib-rs](https://github.com/trifectatechfoundation/zlib-rs),
+whose output was byte-identical to zlib-ng's at levels 1, 6 and 9 on the
+benchmark inputs and which matched its HDF5 read and write speed within 6%
+(tank, 2026-09-23, [BENCHMARKS.md](BENCHMARKS.md#deflate-backend-zlib-rs-vs-zlib-ng)). CI
+fails if a C-building crate enters their default dependency tree. C comes in
+only when you ask: `fast-deflate` (zlib-ng, needs cmake), `zstd`, `szip` (a `clawhdf5-format` feature),
+`https` and the cloud stores (ring / aws-lc-rs), the BLAS backends,
+`clawhdf5-migrate` (bundled SQLite) and the Node.js bindings. One exception
+links rather than builds C: on macOS the default `system-zlib-decompress`
+feature inflates with the system libz first.
+
+## Install
+
+The crates are not on crates.io; depend on the repository (MSRV 1.92):
+
+```toml
+[dependencies]
+clawhdf5        = { git = "https://git.redclaw.dev/quantumclaw/clawhdf5" }
+# optional parts
+clawhdf5-remote = { git = "https://git.redclaw.dev/quantumclaw/clawhdf5" }   # HTTP / object stores
+clawhdf5-agent  = { git = "https://git.redclaw.dev/quantumclaw/clawhdf5" }   # agent memory
 ```
 
-> **New here?** Start with the **[Quickstart Guide](docs/QUICKSTART.md)** · See **[Use Cases](docs/USE_CASES.md)** · Read **[Benchmarks](BENCHMARKS.md)**
+or, with a checkout, `clawhdf5 = { path = "../clawhdf5/crates/clawhdf5" }`.
+Add `features = ["plugin-filters"]` for every plugin filter.
 
----
+The Python package is not on PyPI; build it with
+[maturin](https://www.maturin.rs) into a virtualenv:
 
-## Why ClawhDF5?
-
-Every AI agent needs memory. Today that means scattered Markdown files, SQLite databases, cloud-hosted vector stores, and glue code. ClawhDF5 replaces all of it:
-
-| Problem | Status Quo | ClawhDF5 |
-|---------|-----------|----------|
-| Vector search | External DB (Pinecone, Qdrant) | Built-in, sub-millisecond |
-| Keyword search | Separate FTS engine | Integrated BM25 |
-| Knowledge graph | Neo4j or none | In-file graph with spreading activation |
-| Memory consolidation | Manual pruning | Hippocampal-inspired automatic tiers |
-| Temporal queries | Custom code | Native temporal index (716ns) |
-| Multi-modal | Multiple stores | Unified cross-modal search |
-| Security | Hope for the best | Provenance tracking + anomaly detection |
-| Portability | Config + DB + files | **One `.h5` file. Copy it anywhere.** |
-
----
-
-## Performance
-
-Vector search and agent-memory operations below are benchmarked on Intel i7-12650H (10C/16T), 384-dim embeddings, Criterion.rs. The HDF5 Core I/O table immediately below is from a separate, independently reproduced run (see its own hardware note).
-
-### HDF5 Core I/O (vs libhdf5 1.14.6)
-
-*Benchmark numbers are being validated in collaboration with engineers from the HDF5 Group to confirm methodology and reproducibility.*
-
-Figures below are from an independent reproduction run on a second machine (AMD Ryzen 7 7800X3D, 2026-08-03). Full methodology, the original i7-12650H run, and two additional benchmarks added to close prior coverage gaps (an I/O-inclusive metadata-open comparison and an honest zero-copy-mmap measurement) are in [BENCHMARKS.md § Independent Validation](BENCHMARKS.md#independent-validation-tank-ryzen-7-7800x3d-2026-08-03).
-
-| Operation | ClawhDF5 | libhdf5 | Speedup |
-|-----------|----------|---------|---------|
-| Attribute write (128 attrs) | 85.2 µs | 877 µs | **10.3×** |
-| Group create (64 groups) | 130 µs | 1.37 ms | **10.6×** |
-| Chunked write, deflate-6 (512×512 f32) | 1.44 ms | 65.0 ms | **45.3×** |
-| Sequential read (100K f32) | 23.3 µs | 63.6 µs | **2.7×** |
-| Sequential write (100K f32) | 210 µs | 189 µs | **≈ tie** |
-
-### Vector Search
-
-| Scale | Flat | IVF (nprobe=10) | IVF-PQ | vs MemX¹ |
-|-------|------|-----------------|--------|----------|
-| 1K | **54 µs** | — | — | — |
-| 10K | 753 µs | **27 µs** | — | — |
-| 100K | 11.4 ms | 1.32 ms | **1.19 ms** | ~8–76× (see caveat) |
-
-> Reproduced on the same second machine (Ryzen 7 7800X3D) with a corrected,
-> apples-to-apples SIMD/scalar/parallel comparison methodology — see
-> [BENCHMARKS.md § Independent Validation: tank — LongMemEval & Vector
-> Search](BENCHMARKS.md#independent-validation-tank--longmemeval--vector-search-ryzen-7-7800x3d-2026-08-05).
-
-### Agent Memory Operations
-
-| Operation | Latency | Scale |
-|-----------|---------|-------|
-| Hybrid search (RRF) | **222 µs** | 1K records |
-| BM25 keyword search | **67 µs** | 1K records |
-| Knowledge graph BFS | **24 µs** | 1K entities |
-| Spreading activation | **17 µs** | 100 entities |
-| Temporal range query | **716 ns** | 10K timestamps |
-| Consolidation cycle | **164 µs** | 1K records |
-| Memory write (WAL) | **18 µs** | per record (group-commit append; HDF5 batched at flush) |
-| Importance gate | **61 ns** | per record |
-
-### Chunked Write Throughput (codec comparison)
-
-Measured with Criterion on f32 matrices. Auto-shuffle is applied before all compression codecs
-by default (AoS→SoA byte transpose, +157–204% throughput for float data):
-
-| Codec | 128×128 f32 | 512×512 f32 | Notes |
-|-------|-------------|-------------|-------|
-| Zstd level 3 | **148 µs / 422 MiB/s** | **1.34 ms / 748 MiB/s** | With auto-shuffle |
-| Deflate level 6 | 153 µs / 407 MiB/s | 1.39 ms / 719 MiB/s | With auto-shuffle |
-| Pcodec | 528 µs / 118 MiB/s | 1.69 ms / 591 MiB/s | Best compression ratio |
-
-Use `.with_zstd(3)` or `.with_deflate(6)` for write-heavy workloads — both now perform at ~720–750 MiB/s on large matrices. Use `.with_pcodec()` for write-once/read-many workloads where compression ratio matters more than encode speed. Disable auto-shuffle with `.without_shuffle()` for byte arrays that don't benefit from AoS→SoA transposition.
-
-> ¹ MemX ([arxiv:2603.16171](https://arxiv.org/abs/2603.16171), March 2026): Rust + libSQL, claims <90ms at 100K records. **Not like-for-like:** MemX's figure is *end-to-end* (embeddings + FTS5 + four-factor re-ranking); ours is a *single component* (raw vector search). The ratio overstates the real advantage by an unquantified margin — order-of-magnitude indication only. See [BENCHMARKS.md](BENCHMARKS.md#comparison-to-memx-arxiv260316171).
-
-### LongMemEval Retrieval Recall
-
-Evaluated against the full **`longmemeval_s`** haystack — all 500 questions, 47.7
-sessions and 493.5 turns each, with only 4.0% of haystack sessions being evidence
-sessions. See [BENCHMARKS.md § LongMemEval
-Results](BENCHMARKS.md#longmemeval-results) for the full scoring-target
-declaration:
-
-| Mode | Turn-Level Hit@5 | Session-Level Hit@5 |
-|------|------------------|---------------------|
-| BM25 only | 75.0% | 93.6% |
-| Vector only (MiniLM) | 71.8% | 94.2% |
-| Hybrid (0.4/0.6, tuned) | **81.4%** | **96.8%** |
-
-Hybrid is the strongest configuration, which is what running two retrieval stages
-is for. The weights matter more than the stages: a sweep of `vector_weight` from
-0.0 to 1.0 found the long-standing `0.7/0.3` default is **strictly dominated** by
-`0.4/0.6` — better on Hit@1, Hit@5, Hit@10 and MRR at both granularities. Use
-`0.4/0.6`, or `0.3/0.7` if rank-1 precision matters most. See
-[BENCHMARKS.md § Weight sweep](BENCHMARKS.md#longmemeval-results).
-
-Vector embeddings require `--features embeddings`; without it the vector stage is
-inert and only the BM25 row is produced, which is what every previously published
-number here measured.
-
-On the easier `longmemeval_oracle` variant (evidence sessions only) the same
-harness scores 84.4% turn-level Hit@5 / MRR 0.6597, reproduced identically on a
-second machine. The 9.4-point gap is the cost of the real haystack, and is why the
-full-haystack number is the one quoted here.
-
-This is **retrieval recall** (did the gold memory appear in the top-k), not the
-official LongMemEval QA-accuracy metric — the two are not comparable, and
-retrieval recall reported as QA accuracy typically overstates by 20–30 points.
-
-> **Previously reported here and now retracted:** session-level Hit@5 of 100.0% /
-> MRR 1.0000, and a claim of beating MemX's 51.6%. Those session-level figures were
-> degenerate on the oracle variant (any returned document is a hit by
-> construction); the 93.6% above is a different, real measurement on a corpus where
-> evidence sessions are 4.0% of the haystack. The MemX comparison stays withdrawn —
-> MemX measures fact-level granularity over 220,349 records, which running the full
-> haystack does not fix. Details in
-> [BENCHMARKS.md](BENCHMARKS.md#retracted-session-level-recall-and-the-memx-comparison).
-
-> Enable embeddings via `hybrid_search(query_emb, text, 0.4, 0.6, k)` for substantially higher recall. The vector stage is served by the HNSW index by default (the `hnsw` feature is on by default); build with `--no-default-features --features float16` to fall back to an exact linear cosine scan.
-
-### Memory Footprint
-
-| Records | File Size | Bytes/Record | With Compression |
-|---------|-----------|--------------|------------------|
-| 1K | ~6.5 MB | ~6.5 KB | ~2.1 MB (3.1x) |
-| 10K | ~65 MB | ~6.5 KB | ~21 MB (3.1x) |
-| 100K | ~645 MB | ~6.5 KB | ~208 MB (3.1x) |
-
-### Consolidation Efficiency
-
-| Metric | Before | After | Delta |
-|--------|--------|-------|-------|
-| Records in store | 1,000 | ~110 | −89% |
-| Hit@1 recall | ~60% | ~90% | +30% |
-| Search latency | ~2.8 ms | ~0.3 ms | **9x faster** |
-
-**Full benchmark details: [BENCHMARKS.md](BENCHMARKS.md)**
-
----
-
-## Agent Memory Architecture
-
-ClawhDF5's agent memory engine implements research from 15+ recent papers on agentic memory systems. It's not a toy — it's the real thing.
-
-```
-                        ┌─────────────────┐
-                        │   Agent Query    │
-                        └────────┬────────┘
-                                 │
-                    ┌────────────▼────────────┐
-                    │   Hybrid Retrieval      │
-                    │  Vector + BM25 + RRF    │
-                    └────────────┬────────────┘
-                                 │
-              ┌──────────────────▼──────────────────┐
-              │         Multi-Factor Re-Ranking      │
-              │  temporal · authority · activation    │
-              └──────────────────┬──────────────────┘
-                                 │
-                    ┌────────────▼────────────┐
-                    │  Confidence Rejection   │
-                    │  (suppress bad matches) │
-                    └────────────┬────────────┘
-                                 │
-        ┌────────────────────────▼────────────────────────┐
-        │              Memory Store (HDF5)                │
-        │                                                 │
-        │  ┌───────────┐ ┌───────────┐ ┌───────────────┐  │
-        │  │ Working   │→│ Episodic  │→│  Semantic     │  │
-        │  │ (bounded) │ │ (bounded) │ │ (long-term)   │  │
-        │  └───────────┘ └───────────┘ └───────────────┘  │
-        │                                                 │
-        │  ┌──────────┐ ┌──────────┐ ┌────────────────┐   │
-        │  │Knowledge │ │Temporal  │ │  Multi-Modal   │   │
-        │  │  Graph   │ │  Index   │ │  Embeddings    │   │
-        │  └──────────┘ └──────────┘ └────────────────┘   │
-        │                                                 │
-        │  ┌──────────┐ ┌──────────┐ ┌────────────────┐   │
-        │  │Provenance│ │ Anomaly  │ │   Source       │   │
-        │  │ Tracking │ │Detection │ │  Isolation     │   │
-        │  └──────────┘ └──────────┘ └────────────────┘   │
-        └─────────────────────────────────────────────────┘
-                              │
-                     ┌────────┴────────┐
-                     │ agent_memory.h5 │
-                     │   single file   │
-                     └─────────────────┘
+```bash
+python -m venv .venv && . .venv/bin/activate
+pip install maturin numpy
+maturin develop --release -m crates/clawhdf5-py/Cargo.toml   # add --features https for https://
+python -c "import clawhdf5; print(clawhdf5.__version__)"
 ```
 
-### Module Overview
+`h5rs`: `cargo install --path crates/clawhdf5-tools` (add
+`--features remote` for URLs).
 
-| Module | What It Does |
-|--------|-------------|
-| **`knowledge`** | Entity/relation graph with BFS traversal, spreading activation, fuzzy entity resolution |
-| **`consolidation`** | Three-tier memory (Working → Episodic → Semantic) with importance scoring and time-decay |
-| **`hybrid`** | Vector + BM25 fusion with Reciprocal Rank Fusion (RRF, k=60). The vector stage uses the HNSW index by default (`hnsw` feature, on by default); disable with `--no-default-features --features float16` for an exact linear scan |
-| **`reranker`** | Multi-factor re-ranking: temporal recency, source authority, activation weight |
-| **`confidence`** | Low-confidence rejection — suppresses spurious recalls when nothing matches |
-| **`temporal`** | Sorted timestamp index, session DAG, entity timeline, temporal query hints |
-| **`multimodal`** | Cross-modal search across text/image/audio/video embeddings |
-| **`provenance`** | Source attribution, FNV-1a content hashing, integrity verification |
-| **`anomaly`** | Write rate limiting, 15 injection pattern detectors, source distribution analysis |
-| **`openclaw`** | OpenClaw integration: MemoryBackend trait, Markdown ↔ HDF5 conversion |
-| **`vector_search`** | Flat cosine, pre-normed, SIMD, BLAS, GPU, parallel search paths |
-| **`ivf` / `pq`** | IVF-PQ approximate nearest neighbor for billion-scale search |
-| **`bm25`** | BM25 keyword index with TF-IDF scoring |
-| **`entity_extract`** | Rule-based entity extraction from text chunks into the knowledge graph |
-| **`wal`** | Write-ahead log for crash-safe persistence; each entry is CRC32-checked on replay, so a corrupted entry stops replay there instead of loading bad data |
-| **`memory_strategy`** | Pluggable strategies: save-every, semantic-shift, user-correction detection |
-| **`decision_gate`** | Sub-microsecond trivial/substantive classification |
-| **`async_memory`** | Tokio-based async wrapper over the memory store (`async` feature) |
-
----
-
-## Quick Start
-
-### HDF5 File I/O
+## Quick start: Rust
 
 ```rust
-use clawhdf5::{File, FileBuilder, AttrValue};
+use clawhdf5::{AttrValue, File, FileBuilder, Selection};
 
-// Write
-let mut builder = FileBuilder::new();
-builder.create_dataset("temperatures")
-    .with_f64_data(&[22.5, 23.1, 21.8])
-    .with_shape(&[3]);
-builder.write("output.h5")?;
+// Write: a chunked, deflate-compressed 2-D dataset that can grow along axis 0.
+let data: Vec<f64> = (0..1000 * 64).map(|i| i as f64).collect();
+let mut b = FileBuilder::new();
+b.create_dataset("run/temps")          // intermediate groups are created, as in h5py
+    .with_f64_data(&data)
+    .with_shape(&[1000, 64])
+    .with_maxshape(&[u64::MAX, 64])    // u64::MAX = unlimited
+    .with_chunks(&[100, 64])
+    .with_deflate(6)
+    .set_attr("units", AttrValue::String("K".into()));
+b.write("data.h5")?;
 
-// Read
-let file = File::open("output.h5")?;
-let ds = file.dataset("temperatures")?;
-let values = ds.read_f64()?;
-assert_eq!(values, vec![22.5, 23.1, 21.8]);
-```
-
-### Agent Memory
-
-```rust
-use clawhdf5_agent::{HDF5Memory, MemoryConfig, MemoryEntry, AgentMemory};
-
-// Create memory store
-let config = MemoryConfig::new("agent.h5", "my-agent", 384);
-let mut memory = HDF5Memory::create(config)?;
-
-// Save a memory
-memory.save(MemoryEntry {
-    chunk: "User prefers dark mode and vim keybindings.".into(),
-    embedding: embed("User prefers dark mode..."),  // your embedder
-    source_channel: "chat".into(),
-    timestamp: now(),
-    session_id: "session-001".into(),
-    tags: "preference".into(),
+// Read: whole datasets, or a hyperslab (only the chunks it touches are decoded).
+let file = File::open("data.h5")?;
+let ds = file.dataset("run/temps")?;
+assert_eq!(ds.shape()?, vec![1000, 64]);
+let all = ds.read_f64()?;
+let rows = ds.read_f64_selection(&Selection::Hyperslab {
+    start: vec![10, 0], stride: vec![1, 1], count: vec![2, 64], block: vec![1, 1],
 })?;
+assert_eq!(rows.len(), 128);
+println!("{:?} {:?}", ds.attr("units")?, file.root().groups()?);
+```
 
-// Search
-let results = memory.search(&query_embedding, 5)?;
-for result in results {
-    println!("[{:.3}] {}", result.score, result.chunk);
+Edit that file in place — no rewrite; each call is written and synced
+before it returns, and anything the editor cannot do safely is refused
+before a byte is written:
+
+```rust
+use clawhdf5::{AttrValue, FileEditor, Selection};
+
+let mut ed = FileEditor::open("data.h5")?;          // exclusive lock, as libhdf5 takes
+ed.resize("run/temps", &[1100, 64])?;              // h5py: ds.resize((1100, 64))
+let sel = Selection::Hyperslab {
+    start: vec![1000, 0], stride: vec![1, 1], count: vec![100, 64], block: vec![1, 1],
+};
+ed.write_values("run/temps", &sel, &vec![0.5f64; 100 * 64])?; // ds[1000:1100] = 0.5
+ed.set_attr("run/temps", "calibrated", &AttrValue::I64(1))?;
+```
+
+The editor changes chunk indexes and heaps as libhdf5 does (the tests
+compare index shapes and heap bookkeeping with libhdf5's, and check every
+edited file with h5py, h5dump and `h5rs check`). More in
+[docs/QUICKSTART.md](docs/QUICKSTART.md): groups and links, filters,
+strings and variable-length data, NetCDF-4.
+
+## Quick start: Python
+
+```python
+import numpy as np
+import clawhdf5
+
+with clawhdf5.File("data.h5", "r") as f:
+    print(list(f.keys()))              # member names, like h5py
+    ds = f["group/temperatures"]       # relative or absolute paths
+    print(ds.shape, ds.dtype, ds.chunks)
+    block = ds[100:200, ::4]           # a small selection decodes only its chunks
+    row = ds[-1]                       # integers drop the axis
+    picked = ds[[1, 5, 9], :]          # one increasing index list per key
+    units = ds.attrs["units"]          # attributes come back as h5py returns them
+    everything = np.asarray(ds)
+    ids = f["table"]["id"]             # compound -> structured array; one field
+
+with clawhdf5.File("data.h5", "r+") as f:          # edited in place, h5py semantics
+    f["group/temperatures"][100:200, ::4] = 0.0
+    f["series"].resize(5000, axis=0)   # chunked datasets, within maxshape
+    f["series"][4000:] = np.ones(1000)
+    f["group"].attrs["calibrated"] = True
+
+with clawhdf5.File("http://data.example.org/run42.h5") as f:   # range requests, no download
+    first = f["group/temperatures"][0]
+```
+
+Reads release the GIL, so Python threads read in parallel. The test suite
+(`crates/clawhdf5-py/tests`) compares every read and every edit with h5py,
+locally and over HTTP. Types, keys, writing (`'w'`: numeric arrays) and
+limits: [crates/clawhdf5-py/README.md](crates/clawhdf5-py/README.md).
+
+## Remote files, the browser, SWMR
+
+**Remote files** ([`clawhdf5-remote`](crates/clawhdf5-remote/README.md),
+design: [docs/design/range-reads.md](docs/design/range-reads.md)). The
+same `clawhdf5::File`, over HTTP range requests or an object store, through
+a block cache (1 MiB blocks, LRU budget, concurrent requests deduplicated,
+runs coalesced into parallel requests). The file is pinned by ETag /
+Last-Modified and length: a file that changes on the server is an error,
+never a mix of old and new bytes.
+
+```rust
+let file = clawhdf5_remote::open_url("http://127.0.0.1:8000/tall.h5")?;
+let values = file.dataset("/g2/dset2.1")?.read_f64()?;
+```
+
+Try it with the crate's test server:
+
+```bash
+cargo run -p clawhdf5-remote --example range_server -- crates/clawhdf5/tests/fixtures 127.0.0.1:8000
+cargo run -p clawhdf5-remote --example read_url -- http://127.0.0.1:8000/tall.h5 /g2/dset2.1
+```
+
+Plain HTTP builds no C; `https` (rustls + ring) and `s3` / `gcs` / `azure`
+are opt-in features. The cloud backends are built and their URL handling
+tested, but have not been run against a real bucket.
+
+**In the browser** (`clawhdf5-wasm`, demo and API in
+[examples/wasm-viewer](examples/wasm-viewer/README.md)): `open(bytes)` reads
+a file held in memory; `openUrl(url)` reads a file on a web server by range
+requests, fetching only what each call needs, on the main thread (no
+worker, no synchronous XHR). In a 200 MB h5py file, listing the root,
+reading two small datasets, a group's attributes, the large dataset's shape
+and a 10-value window of it took 5 requests and 6 MiB (tank, 2026-09-27,
+the viewer's Node + Chromium test suite).
+
+**SWMR reading** (design: [docs/design/swmr.md](docs/design/swmr.md)).
+`File::open_swmr` follows a file a libhdf5 SWMR writer (h5py
+`f.swmr_mode = True`) is still appending to, as h5py's
+`File(path, "r", swmr=True)` does: `Dataset::refresh()` picks up the new
+extent, and a read that races the writer (a checksum failing mid-flush) is
+retried, up to 100 attempts as in libhdf5, never returned torn. Tested live
+against an h5py writer appending to Extensible-Array and v2-B-tree indexed
+datasets for 2 500 steps (20 000 in a release build), beside h5py's own
+SWMR reader. clawhdf5 does not write SWMR files.
+
+```rust
+let file = clawhdf5::File::open_swmr("live.h5")?;
+let mut ds = file.dataset("samples")?;
+while file.swmr_writer_active()? {   // add your own timeout: a writer that died keeps the flag set
+    ds.refresh()?;
+    let n = ds.shape()?[0];
+    // read the new rows ...
+    std::thread::sleep(std::time::Duration::from_millis(100));
 }
 ```
 
-### Knowledge Graph
+## h5rs tools
 
-```rust
-use clawhdf5_agent::knowledge::KnowledgeCache;
-
-let mut kg = KnowledgeCache::new();
-
-// Add entities
-let alice = kg.add_entity("Alice", "person", -1);
-let bob = kg.add_entity("Bob", "person", -1);
-let acme = kg.add_entity("Acme Corp", "company", -1);
-
-// Add relations
-kg.add_relation(alice, acme, "works_at", 1.0);
-kg.add_relation(bob, acme, "works_at", 1.0);
-kg.add_relation(alice, bob, "manages", 0.8);
-
-// Traverse
-let neighbors = kg.bfs_neighbors(alice, 2);  // 2-hop neighborhood
-
-// Spreading activation — find related entities
-let activated = kg.spreading_activation(&[alice], 0.5, 0.01, 5);
-
-// Entity resolution — fuzzy matching
-let resolved = kg.resolve_or_create("alice", "person", -1, 2);
-// Returns existing Alice entity (Levenshtein distance ≤ 2)
-```
-
-### Memory Consolidation
-
-```rust
-use clawhdf5_agent::consolidation::*;
-
-let config = ConsolidationConfig::default();
-let mut engine = ConsolidationEngine::new(config);
-
-// Add memories — automatically scored for importance
-engine.add_memory("User prefers dark mode", vec![0.1, 0.2, ...], MemorySource::User);
-engine.add_memory("ok", vec![0.0, 0.0, ...], MemorySource::System);
-
-// Access a memory (reactivates it)
-engine.access_memory(0);
-
-// Run consolidation cycle
-let stats = engine.consolidate();
-// Working memories promote to Episodic (if important enough)
-// Episodic memories promote to Semantic (if accessed enough)
-// Low-decay memories get evicted when tiers are full
-```
-
-### Temporal Queries
-
-```rust
-use clawhdf5_agent::temporal::*;
-
-let mut index = TemporalIndex::new();
-index.insert(1, 1700000000.0);  // record 1 at timestamp
-index.insert(2, 1700003600.0);  // record 2, 1 hour later
-
-// Range query — "what happened between 2pm and 5pm?"
-let ids = index.range_query(1700000000.0, 1700010800.0);
-
-// Latest 10 memories
-let recent = index.latest(10);
-```
-
-### OpenClaw Integration
-
-```rust
-use clawhdf5_agent::openclaw::*;
-
-// Create backend
-let mut backend = ClawhdfBackend::create("memory.h5", "agent-1", 384)?;
-
-// Ingest existing Markdown memory files
-let md = std::fs::read_to_string("MEMORY.md")?;
-let count = backend.ingest_markdown("MEMORY.md", &md)?;
-
-// Search (uses full pipeline: RRF → re-rank → confidence filter)
-let results = backend.search("user preferences", &query_embedding, 5);
-
-// Export back to Markdown
-let exported = backend.export_markdown("MEMORY.md")?;
-```
-
----
-
-## Crate Map
-
-```
-clawhdf5 workspace (16 crates, ~92K lines of Rust; plus libaec-sys, an
-                     internal FFI bindings crate for the optional szip feature)
-│
-├── Core HDF5
-│   ├── clawhdf5-format      — Binary parser/writer (no_std), shared type definitions
-│   ├── clawhdf5-io          — I/O abstraction (buffered, mmap, async)
-│   ├── clawhdf5-filters     — Fast deflate path (zlib-ng); lz4/zstd/pcodec/szip filters live in clawhdf5-format
-│   ├── clawhdf5-derive      — Proc macros
-│   ├── clawhdf5             — High-level API
-│   ├── clawhdf5-netcdf4     — NetCDF-4 support
-│   ├── clawhdf5-accel       — SIMD (NEON, AVX2, AVX-512)
-│   └── clawhdf5-gpu         — GPU compute (wgpu, hand-written WGSL compute shaders)
-│
-├── Agent Memory
-│   ├── clawhdf5-agent       — Memory engine (20.9K lines, 32 modules; WAL is CRC32-checked per entry)
-│   ├── clawhdf5-ann         — HNSW approximate nearest neighbor (default backend; optional `parallel` feature)
-│   ├── clawhdf5-migrate     — SQLite → HDF5 migration
-│   ├── clawhdf5-android     — Android JNI bridge
-│   └── clawhdf5-cli         — CLI tool
-│
-├── Bindings
-│   ├── clawhdf5-py          — Python (PyO3)
-│   └── clawhdf5-napi        — Node.js (napi-rs)
-│
-└── Tooling
-    └── clawhdf5-bench       — Benchmark suite
-```
-
----
-
-## Research Foundation
-
-ClawhDF5's agent memory design draws from 15+ recent papers:
-
-| Paper | Key Insight | ClawhDF5 Module |
-|-------|-------------|-----------------|
-| **MemX** (2026) | RRF + multi-factor re-ranking | `hybrid`, `reranker` |
-| **Graph-Native Cognitive Memory** (2026) | Graph-structured belief revision | `knowledge` |
-| **CraniMem** (2026) | Bounded hippocampal memory | `consolidation` |
-| **D-MEM** (2026) | Reward prediction error gating | `consolidation` |
-| **SYNAPSE** (2025) | Spreading activation for recall | `knowledge` |
-| **RAGdb** (2025) | Zero-dependency edge RAG | Architecture |
-| **MemoryGraft** (2025) | Memory poisoning attacks | `anomaly`, `provenance` |
-| **MemoryArena** (2026) | Multi-session benchmark | `temporal` |
-| **AI Hippocampus** (2026) | Memory taxonomy survey | Overall design |
-
----
-
-## Feature Flags
-
-### `clawhdf5-agent`
-
-| Flag | Default | Description |
-|------|---------|-------------|
-| `agent` | no | Full agent memory layer |
-| `float16` | **yes** | Half-precision embedding storage (2× compression) |
-| `hnsw` | **yes** | HNSW approximate vector index for `hybrid_search` (via `clawhdf5-ann`); disable for an exact linear scan |
-| `parallel` | no | Rayon parallel search |
-| `fast-math` | no | BLAS matrix-vector multiply |
-| `accelerate` | no | Apple Accelerate / AMX (macOS) |
-| `openblas` | no | OpenBLAS (Linux) |
-| `gpu` | no | GPU search via wgpu |
-| `async` | no | Tokio async with background flush |
-
-### `clawhdf5-format`
-
-| Flag | Default | Description |
-|------|---------|-------------|
-| `std` | yes | Standard library (disable for `no_std`) |
-| `deflate` | yes | Deflate compression |
-| `checksum` | yes | Jenkins lookup3 verification |
-| `provenance` | yes | SHA-256 provenance attributes |
-| `fast-deflate` | **yes** | zlib-ng backend for faster deflate |
-| `system-zlib-decompress` | **yes** | Use the system zlib for decompression where available |
-| `parallel` | no | Parallel chunk encoding + compression (rayon) |
-| `fast-checksum` | no | crc32fast-accelerated checksums |
-| `lz4` | no | LZ4 block compression filter (id 32004) |
-| `zstd` | no | Zstandard compression filter (id 32015) |
-| `pcodec` | no | Pcodec lossless numerical codec (id 32023, via `pco` crate) |
-| `system-zlib` / `zlib-rs` | no | Alternative zlib backends for deflate |
-| `blake3_hash` | no | BLAKE3 content hashing for provenance |
-
-### `clawhdf5-ann`
-
-| Flag | Default | Description |
-|------|---------|-------------|
-| `parallel` | no | Rayon-parallel neighbor-distance computation during HNSW graph pruning |
-
-### `clawhdf5-io`
-
-| Flag | Default | Description |
-|------|---------|-------------|
-| `mpi-io` | no | MPI-backed I/O via the `mpi` crate |
-
-> **Parallel I/O (MPI) limitation:** `mpi-io`'s read path is a root-rank read
-> followed by a broadcast, and its write path gathers all ranks' shards to
-> rank 0 before writing — not true collective I/O
-> (`MPI_File_read_at_all`/`write_at_all`). It does not provide I/O bandwidth
-> that scales with rank count; true collective I/O is tracked as future work.
-
----
-
-## Building
+`h5rs` (crate `clawhdf5-tools`) is a pure-Rust counterpart of the HDF5
+command-line tools:
 
 ```bash
-# Default
-cargo build --workspace
-
-# Agent memory with all accelerations (Linux)
-cargo build -p clawhdf5-agent --features "agent,float16,parallel,fast-math"
-
-# Agent memory with Apple Accelerate (macOS)
-cargo build -p clawhdf5-agent --features "agent,float16,accelerate,parallel,gpu"
-
-# Tests
-cargo test --workspace            # all 1,650+ tests
-cargo test -p clawhdf5-agent      # agent memory tests
-
-# Benchmarks
-cargo bench -p clawhdf5-agent               # agent memory suite
-cargo bench -p clawhdf5-bench               # h5bench-equivalent I/O suite
+h5rs ls -r file.h5          # like h5ls
+h5rs dump file.h5           # like h5dump: DDL, or --json (hdf5-json)
+h5rs stat file.h5           # like h5stat
+h5rs diff a.h5 b.h5         # like h5diff
+h5rs check --data file.h5   # structural and checksum validator
 ```
 
----
+`dump` output is byte-identical to h5dump's on the interop test files, and
+the `ls`/`stat`/`diff` tests compare with h5ls, h5stat and h5diff. `check`
+walks the file's structures, verifies their checksums (superblock, object
+headers, v2 B-trees, fractal heaps, chunk indexes) and with `--data`
+decodes every dataset; it validates with the library's own parsers, so it
+accepts what they accept. With `--features remote` every subcommand takes a
+URL. Details: [crates/clawhdf5-tools/README.md](crates/clawhdf5-tools/README.md).
 
-## HDF5 File Schema
+## Agent memory
 
-```
-agent_memory.h5
-├── /meta
-│   ├── schema_version: "1.0"
-│   ├── agent_id, embedder, embedding_dim
-│   └── created_at
-├── /memory
-│   ├── chunks:      string[N]
-│   ├── embeddings:  f32[N × D]  (or f16 with float16 flag)
-│   ├── tombstones:  u8[N]
-│   └── norms:       f32[N]      (pre-computed L2)
-├── /sessions
-│   ├── ids:         string[S]
-│   └── summaries:   string[S]
-└── /knowledge_graph
-    ├── entity_names:    string[E]
-    ├── relation_srcs:   i64[R]
-    ├── relation_tgts:   i64[R]
-    └── relation_types:  string[R]
+`clawhdf5-agent` stores an agent's memories — text, embeddings, sessions,
+a knowledge graph — in one HDF5 file (readable by h5py), with:
+
+- **Hybrid search**: HNSW (clawhdf5-ann) vector + BM25 keyword, weighted
+  0.4 / 0.6, optional source filter, re-ranking and confidence rejection.
+  On the full LongMemEval `longmemeval_s` haystack (500 questions, real
+  MiniLM embeddings) turn-level Hit@5 is **81.4%** — retrieval recall, not
+  the official QA-accuracy metric (tank, re-run 2026-09-27,
+  [BENCHMARKS.md](BENCHMARKS.md#longmemeval-results)).
+- **Compact by default**: float16 embeddings on disk (48% smaller at 100K;
+  tank, 2026-09-23) and an int8 index copy with exact re-scoring — 1.74x
+  the raw vectors in memory at 100K instead of 2.72x, and 1.63x the QPS at
+  equal recall on AVX2 (int8 side measured 2026-09-19/20, machine not
+  recorded, not re-run since; see
+  [BENCHMARKS.md](BENCHMARKS.md#quantising-the-index-copy-quantized_index)).
+- **Durability**: a write-ahead log with a chained CRC per entry,
+  crash-safe checkpoints, a single-writer lock and a read-only open. WAL
+  appends are not fsynced: saves since the last checkpoint can be lost on
+  power failure.
+- **Signed checkpoints**: Ed25519 over a SHA-256 Merkle tree of the
+  records, settings, sessions and graph; `HDF5Memory::verify` names the
+  edited records.
+
+```rust
+use clawhdf5_agent::{AgentMemory, HDF5Memory, MemoryConfig, MemoryEntry, SearchOptions};
+
+let mut memory = HDF5Memory::create(MemoryConfig::new("agent.h5".into(), "my-agent", 384))?;
+memory.save(MemoryEntry {
+    chunk: "User prefers dark mode and vim keybindings.".into(),
+    embedding: embed("User prefers dark mode and vim keybindings."), // your embedder
+    source_channel: "chat".into(),
+    timestamp: now,
+    session_id: "session-001".into(),
+    tags: "preference".into(),
+})?;
+for r in memory.search(&embed("what editor?"), "editor preferences", &SearchOptions::new(5)) {
+    println!("[{:.3}] {}", r.score, r.chunk);
+}
 ```
 
----
+Architecture, every module, performance tables, feature flags, file
+schema, CLI and SQLite migration: [docs/agent-memory.md](docs/agent-memory.md).
 
-## Migration
+## Crate map
 
-### From rustyhdf5 / edgehdf5
+19 crates under `crates/`, plus `libaec-sys` (FFI for the optional SZIP
+filter).
 
-Replace in `Cargo.toml` and source:
+| Crate | Role |
+|---|---|
+| **HDF5** | |
+| `clawhdf5` | The facade: `File`, `FileBuilder`, `FileEditor`, `Dataset`, `Group`, SWMR reading |
+| `clawhdf5-format` | The format itself (superblock, headers, B-trees, heaps, datatypes), the filter pipeline and registry, every codec but the deflate backends; `no_std`-capable |
+| `clawhdf5-filters` | Deflate backends (zlib-rs default, zlib-ng, Apple Compression) |
+| `clawhdf5-io` | I/O helpers: mmap, async, an HSDS client, `mpi-io` (not collective I/O) |
+| `clawhdf5-remote` | HTTP(S) and object-store files through a block cache |
+| `clawhdf5-netcdf4` | NetCDF-4 dimensions, variables, CF attributes |
+| `clawhdf5-derive` | Derive macros for HDF5-serialisable structs |
+| `clawhdf5-tools` | `h5rs`: `ls`, `dump`, `stat`, `diff`, `check` |
+| **Bindings** | |
+| `clawhdf5-py` | Python (PyO3 + numpy) |
+| `clawhdf5-wasm` | Browser (wasm-bindgen), read-only |
+| `clawhdf5-napi` | Node.js (unpublished; does not work, see known-issues) |
+| `clawhdf5-android` | Android JNI bindings for the agent store |
+| **Agent memory** | |
+| `clawhdf5-agent` | The memory store |
+| `clawhdf5-ann` | HNSW index (`f32` or `i8` storage) |
+| `clawhdf5-accel` | SIMD kernels (AVX2, NEON incl. `SDOT`; AVX-512 behind a feature) |
+| `clawhdf5-gpu` | Vector distance computation on the GPU (wgpu, WGSL); HDF5 I/O is CPU-only |
+| `clawhdf5-migrate` | SQLite → agent store migration |
+| `clawhdf5-cli` | The `clawhdf5` agent-memory CLI |
+| `clawhdf5-bench` | Benchmarks and harnesses |
 
-| Old | New |
-|-----|-----|
-| `rustyhdf5*` | `clawhdf5*` |
-| `edgehdf5-memory` | `clawhdf5-agent` |
-| `edgehdf5` (CLI) | `clawhdf5-cli` |
-
-### From SQLite
+## Building and testing
 
 ```bash
-cargo install --path crates/clawhdf5-migrate
-clawhdf5-migrate --sqlite old.db --hdf5 memory.h5 --agent-id my-agent --embedding-dim 384
+cargo build --workspace              # pure Rust: no cmake or C compiler needed
+cargo test --workspace
+scripts/ci-test.sh                   # what CI runs: fmt, clippy matrix, tests, interop, no_std, no-C check
+conformance/run.sh                   # the conformance report (needs h5py, hdf5plugin, h5dump)
 ```
 
----
+The interop suites need a Python with h5py (and netCDF4, xarray; the
+NetCDF-4 tests of h5netcdf-written files skip without h5netcdf); on a
+PEP 668 system that has to be a virtualenv, which `ci-test.sh` finds as
+`.venv` or through `CLAWHDF5_PYTHON`. Without one they skip; set
+`CLAWHDF5_REQUIRE_INTEROP=1` to make that a failure, as CI does:
 
-## Roadmap
+```bash
+python3 -m venv .venv && .venv/bin/pip install h5py numpy netCDF4 xarray
+```
 
-See [ROADMAP.md](ROADMAP.md) for the full implementation tracker.
+CI (`.gitea/workflows/`) runs `ci-test.sh` on x86-64, lints and tests the
+NEON code on aarch64, and runs the conformance corpus nightly.
 
-**Phase 1 complete** — all 8 tracks delivered:
-- ✅ Knowledge Graph with spreading activation
-- ✅ Hippocampal memory consolidation
-- ✅ RRF hybrid retrieval + re-ranking + confidence rejection
-- ✅ Temporal reasoning with sub-µs queries
-- ✅ Memory security + anomaly detection
-- ✅ Multi-modal memory (text/image/audio/video)
-- ✅ OpenClaw integration layer
-- ✅ Comprehensive Criterion benchmarks
+## Documentation
 
-**Phase 2** — MemoryArena and LongMemEval academic benchmarks are done (see [BENCHMARKS.md](BENCHMARKS.md), reproduced on a second machine); remaining: publish the OpenClaw TypeScript bridge to npm, crates.io/PyPI publishing.
+| | |
+|---|---|
+| [docs/QUICKSTART.md](docs/QUICKSTART.md) | Longer quick starts: HDF5 in Rust and Python, NetCDF-4, agent memory, CLI |
+| [docs/USE_CASES.md](docs/USE_CASES.md) | Where clawhdf5 fits, and where it does not |
+| [docs/agent-memory.md](docs/agent-memory.md) | The agent-memory store in full |
+| [CONFORMANCE.md](CONFORMANCE.md) | The conformance report, generated by `conformance/run.sh` |
+| [BENCHMARKS.md](BENCHMARKS.md) | Every measurement, with date, machine and command |
+| [docs/known-issues.md](docs/known-issues.md) | Open limits and fixed bugs, with dates |
+| [CHANGELOG.md](CHANGELOG.md) | Changes, including everything since v2.7.0 |
+| [docs/README.md](docs/README.md) | Index of every document |
 
----
+## Who uses it
 
-## Part of the RedClaw Ecosystem
-
-ClawhDF5 powers the `.brain` format for [ClawBrainHub](https://clawbrainhub.com) — the brain registry for AI agents. One file that packages identity, skills, memory, knowledge, and cryptographic provenance.
-
----
+[ClawBrainHub](https://clawbrainhub.com) is the one verified consumer: its
+`.brain` files are HDF5 files it reads and writes through the facade
+(`File`, `FileBuilder`, `AttrValue`, `Selection`), and its CLI uses
+`clawhdf5_agent::bm25::BM25Index` (builds and passes its tests against
+`main`, checked 2026-09-25). clawhdf5 is **not** an OpenClaw memory plugin
+([docs/openclaw.md](docs/openclaw.md)), and ZeroClaw does not use it.
 
 ## License
 
-MIT
-
----
-
-<p align="center">
-  <em>Built by <a href="https://github.com/redclawsystems">RedClaw Systems</a></em><br>
-  <em>~92,000 lines of Rust. Zero C dependencies. One file to remember everything.</em>
-</p>
+MIT — see [LICENSE](LICENSE).

ROADMAP.md

@@ -1,187 +1,160 @@
-# ClawhDF5 Roadmap — Agent Memory Evolution
+# clawhdf5 roadmap
 
-> Making clawhdf5 the defacto agentic memory solution.
-> Single file. Pure Rust. Zero dependencies. Trusted everywhere.
+What has shipped, and what is genuinely next. Everything here is checked
+against `CHANGELOG.md`, `git log` and [`docs/known-issues.md`](docs/known-issues.md);
+dates are merge dates on `main`. Nothing after v2.7.0 has been released:
+the work since then is on `main` under `CHANGELOG.md` "Unreleased".
+
+_Last updated: 2026-09-28 (at `9b5803f`, PR #21)._
 
 ---
 
-## Track 1: Knowledge Graph in HDF5
-**Status:** 🟢 Phase 1 Complete
-**Priority:** Critical
-**Crate:** `clawhdf5-agent`
+## Done
 
-- [x] **1.1** Entity storage — entities with properties, embeddings, timestamps (created_at/updated_at)
-- [x] **1.2** Relation storage — typed edges with RelationType enum (Temporal/Causal/Associative/Hierarchical/Custom), metadata, timestamps
-- [x] **1.3** Entity extraction helpers — rule-based extraction (Person, Org, Location, Date, Technology, Project) with extract_and_store_entities() integration
-- [x] **1.4** Entity resolution — fuzzy name matching (Levenshtein distance) via resolve_or_create()
-- [x] **1.5** Graph traversal queries — BFS neighbors with depth, subgraph extraction from seeds
-- [x] **1.6** Spreading activation — weighted activation propagation with configurable decay
-- [x] **1.7** Graph-aware retrieval — get_entity_context() for formatted context injection
-- [x] **1.8** Tests — comprehensive tests for all new features
+### Releases
 
-**Research:** Graph-Native Cognitive Memory (2026), Graph-based Agent Memory survey (2026), SYNAPSE (2025)
+| Version | Date | Headline |
+|---|---|---|
+| v2.0.0 | 2026-03-19 | rustyhdf5 (11 crates) and edgehdf5 (4 crates) unified into one workspace as `clawhdf5-*` |
+| v2.1.0 | 2026-06-03 | HNSW backs the agent's vector search by default; live, mutable HNSW index |
+| v2.2.0 – v2.7.0 | 2026-09-18 – 2026-09-20 | bounded decompression and read-path bounds checks, single-writer store locking, WAL v4, HNSW recall fix (0.31 -> 0.98 recall@10 at 100K), fusion weights tuned on LongMemEval, int8 index, Extensible Array read fix and chunk-index checksums |
+
+Details per release: [`CHANGELOG.md`](CHANGELOG.md).
+
+### Since v2.7.0 (unreleased, on `main`)
+
+| PR | Merged | What |
+|---|---|---|
+| #3 | 2026-09-23 | pure-Rust deflate (zlib-rs) by default, no C in the core crates' default build (checked in CI), MSRV 1.92 |
+| #4 | 2026-09-25 | files open in h5py again (every `f32` and every empty dataset clawhdf5 wrote was unreadable by libhdf5); float16 embedding storage |
+| #5 | 2026-09-25 | `HDF5Memory::search` with `SearchOptions` (source filters, re-ranking, confidence); float16 on by default |
+| #6 | 2026-09-25 | `clawhdf5-migrate` writes real agent stores; knowledge-graph fix; dated benchmark re-run |
+| #7 | 2026-09-25 | consolidation benchmark completed (cheaper novelty scoring) |
+| #8 | 2026-09-25 | Ed25519-signed checkpoints (`HDF5Memory::verify`) |
+| #9, #10 | 2026-09-25 | OpenClaw and ZeroClaw integration claims withdrawn — neither ever integrated clawhdf5 |
+| #11 | 2026-09-26 | silent wrong data and libhdf5 interop bugs found by the HDF5 audit fixed |
+| #12 | 2026-09-26 | reproducible conformance sweep over eight public corpora, nightly CI job ([`CONFORMANCE.md`](CONFORMANCE.md)) |
+| #13 | 2026-09-26 | reads HDF5 1.6-era layouts, user blocks, virtual datasets, dense attributes, very large groups |
+| #14 | 2026-09-26 | `h5rs` tools (`ls`, `dump`, `stat`, `diff`, `check`), the browser reader (`clawhdf5-wasm`), libhdf5's header checks, plugin filters (LZF, bitshuffle, bzip2, Blosc), concurrency benchmark |
+| #15 | 2026-09-26 | fast contiguous and concurrent reads, variable-length data, nested groups and links in the writer, Python bindings |
+| #16 | 2026-09-26 | chunked full reads faster than an h5py process pool, writer B-trees of any size, Blosc2 (read), 599/697 conformance |
+| #17 | 2026-09-26 | range reads M0/M1 (indexed name lookups, the `Storage` trait), ZFP (read), in-place editing (`FileEditor`) |
+| #18 | 2026-09-27 | range reads M2/M3 (`File::open_storage`; `clawhdf5-remote`: HTTP(S), S3, GCS, Azure), in-place editing of every chunk index, shrinking, dense attributes |
+| #19 | 2026-09-27 | remote files in the browser (`openUrl`, M4), SWMR reader (`File::open_swmr`, M5), Python remote reads and `'r+'` editing |
+| #20 | 2026-09-28 | benchmarks re-measured: LongMemEval with real MiniLM embeddings, local reads on an idle machine |
+| #21 | 2026-09-28 | remote files open in a few requests (group lookups down the B-tree, `Storage::hint`), `ObjectHeader::parse` back to its earlier speed, the last conformance mismatches resolved: 602/697 ok, 0 mismatch (the run of 2026-09-28 in [`CONFORMANCE.md`](CONFORMANCE.md) still counts 1 our-error, a corrupt N-Bit file libhdf5's own tests refuse) |
+
+### Range reads (design: [`docs/design/range-reads.md`](docs/design/range-reads.md))
+
+- [x] M0 — indexed name lookups (#17)
+- [x] M1 — metadata parsed through the `Storage` trait (#17)
+- [x] M2 — raw data through `Storage`, `File::open_storage` (#18)
+- [x] M3 — `clawhdf5-remote`: HTTP(S) range requests and object stores through a block cache; `h5rs` URLs (#18); Python URLs (#19)
+- [x] M4 — `openUrl` in the browser, restartable "NeedBytes" cache (#19; fewer round trips in #21)
+- [x] M5 — reading files a SWMR writer is appending to ([`docs/design/swmr.md`](docs/design/swmr.md), #19)
+
+### Agent memory (`clawhdf5-agent`)
+
+Shipped before and during the v2 releases, and kept current since:
+knowledge graph with entity extraction and resolution; three-tier
+consolidation with decay; hybrid retrieval (HNSW + BM25, weighted or RRF
+fusion, re-ranking, confidence rejection, query expansion); temporal index
+and session DAG; per-save provenance ledger and write-anomaly detection;
+multi-modal embeddings; WAL with chained CRC32; single-writer locking;
+signed checkpoints. Retrieval is measured, not claimed: see
+[`BENCHMARKS.md`](BENCHMARKS.md) ("LongMemEval Results" reports retrieval
+recall, not QA accuracy; earlier headline numbers that compared different
+granularities were retracted there).
 
 ---
 
-## Track 2: Memory Consolidation Engine
-**Status:** 🟢 Phase 1 Complete
-**Priority:** Critical
-**Crate:** `clawhdf5-agent`
+## Next
 
-- [x] **2.1** Importance scoring — surprise (novelty), correction boost, length scoring with configurable weights
-- [x] **2.2** Three-tier memory model — Working → Episodic → Semantic with bounded capacities
-- [x] **2.3** Time-decay with reactivation — exponential decay with configurable half-life, access resets timestamp
-- [x] **2.4** Bounded memory with graceful degradation — evict lowest-decay entries when over capacity
-- [x] **2.5** Consolidation cycles — promote/evict across tiers based on importance and access thresholds
-- [x] **2.6** Memory statistics — ConsolidationStats with per-tier counts, eviction/promotion tracking
-- [x] **2.7** Tests — comprehensive tests for all features
+Not scheduled; listed roughly by how much they unblock. None has a date.
 
-**Research:** CraniMem (2026), D-MEM (2026), AI Hippocampus survey (2026)
+### Distribution
+
+- [ ] **Publish the crates to crates.io.** Nothing is published; the READMEs
+  say to depend on git. Before publishing: no `publish` settings exist
+  (only `clawhdf5-wasm` has `publish = false`).
+- [ ] **Publish Python wheels to PyPI.** `crates/clawhdf5-py` builds with
+  maturin and is tested in CI, but no wheel is published. The default wheel
+  reads plain `http://` only; `https`/`s3`/`gcs`/`azure` wheels compile C
+  (ring, aws-lc-rs).
+- [ ] **The Node.js package** (`packages/clawhdf5-node` over
+  `clawhdf5-napi`) has never worked and is not in CI: fix it and add CI, or
+  remove it ([known issue](docs/known-issues.md)).
+
+### HDF5 features
+
+- [ ] **SWMR writing.** The reader is done (M5); writing a file while
+  libhdf5 readers follow it is not. Also not covered: remote SWMR (a remote
+  file is pinned at open), `MmapFile`/`LazyFile` SWMR reads, refreshing
+  groups or attributes.
+- [ ] **MPI collective I/O.** `clawhdf5-io`'s `MpiVol` (`mpi-io`) is
+  root-read + broadcast and gather-to-root writes, not collective MPI-IO
+  (`MPI_File_read_at_all`/`write_at_all`).
+- [ ] **Paged-metadata single-request reads.** Files written with paged
+  aggregation (`H5Pset_file_space_strategy(PAGE)`, `h5repack -S PAGE`)
+  keep their metadata in a few pages; range reads could fetch those in one
+  request and use the file's page size as the block size. Today the block
+  size is fixed (1 MiB) and only the first block is read ahead
+  (range-reads design, option (c) as a policy).
+- [ ] **Blosc2 and ZFP encoders.** Both filters are read-only; the other
+  plugin filters (LZF, bitshuffle, bzip2, Blosc 1) read and write.
+- [ ] **External links and external raw data** are explicit errors, not
+  followed.
+- [ ] **Virtual datasets:** the "first missing" view and printf gaps other
+  than 0, source-to-virtual type conversion other than a byte swap, nested
+  virtual sources, source files outside the virtual file's directory.
+- [ ] **Datatypes:** x87 long double and binary128 are refused.
+- [ ] **Writer:** one attribute or link message over 65 515 bytes in dense
+  storage is an error (huge fractal-heap objects); no option to write
+  files HDF5 1.8 can read.
+- [ ] **`FileEditor`:** new chunks in implicit indexes, variable-length and
+  reference data, filters it cannot encode (scale-offset, N-Bit, SZIP),
+  some dense-attribute heap layouts, creating or deleting objects and
+  attributes (also from Python `'r+'`), and no journal (a crash mid-edit
+  can leave the file inconsistent). Freed space is reused only within one
+  editor.
+- [ ] **Selection reads** decode the whole dataset when the selection's
+  bounding box covers more than half of it (a strided `ds[::100]`), and
+  for compact/virtual datasets or a non-default fill value: correct, but
+  more work than needed.
+- [ ] **Readers:** `LazyFile` and `MmapFile` still need the whole file;
+  the zero-copy methods need the file in memory.
+
+### Remote and browser
+
+- [ ] Run the `s3`/`gcs`/`azure` backends against real buckets (only built
+  and URL-parsing-tested so far).
+- [ ] `h5rs` options for request headers and cache settings.
+- [ ] Browser limits in [`docs/known-issues.md`](docs/known-issues.md)
+  ("`clawhdf5-wasm` (browser) limits"): files of 4 GiB or more (wasm32),
+  compound/reference/opaque datasets, round trips per index level. The
+  package doubled in size with `openUrl`
+  ([size table](examples/wasm-viewer/README.md#size)); dropping the
+  function-name section would take a third off the raw size (13% gzipped).
+
+### Quality
+
+- [ ] Scheduled fuzz campaigns: the cargo-fuzz targets
+  ([`crates/clawhdf5-format/fuzz`](crates/clawhdf5-format/fuzz/README.md),
+  and the agent's WAL target) run only by hand or with
+  `CLAWHDF5_FUZZ_SECONDS`.
 
 ---
 
-## Track 3: Hybrid Retrieval Pipeline
-**Status:** 🟢 Phase 1 Complete
-**Priority:** High
-**Crate:** `clawhdf5-agent`
+## Withdrawn
 
-- [x] **3.1** Reciprocal Rank Fusion (RRF) — rrf_hybrid_search() with k=60 constant
-- [x] **3.2** Multi-factor re-ranking — temporal decay, source authority hierarchy, activation scores (reranker.rs)
-- [x] **3.3** Low-confidence rejection — min_score threshold, gap filtering, max_results (confidence.rs)
-- [x] **3.4** Query expansion — synonyms, acronyms, temporal rewrites, morphological variants, knowledge graph aliases + expanded_search() with RRF merge
-- [x] **3.5** Result explanation — ReRankResult with full score breakdown per factor
-- [x] **3.6** Configurable pipeline — ReRankConfig + ConfidenceConfig with tunable weights/thresholds
-- [x] **3.7** Tests + MemX-comparable benchmarks — 5 integration tests (Hit@1≥90%, search<500ms@100K, BM25<200ms@100K, hybrid<50ms@10K, compact<200ms@10K)
+- **OpenClaw integration** (withdrawn 2026-09-25, PR #9). clawhdf5 was
+  never an OpenClaw memory plugin; the documented
+  `memory.backend = "clawhdf5"` was never valid. The Rust `ClawhdfBackend`
+  remains as a library API. [`docs/openclaw.md`](docs/openclaw.md) records
+  what a real plugin would need.
+- **ZeroClaw integration** (withdrawn 2026-09-25, PR #10). ZeroClaw has no
+  clawhdf5 backend, and `clawhdf5-migrate`'s SQLite layout is not
+  ZeroClaw's schema.
 
-**Research:** MemX (2026), SwiftMem (2026)
-
----
-
-## Track 4: Temporal Reasoning
-**Status:** 🟢 Phase 1 Complete
-**Priority:** High
-**Crate:** `clawhdf5-agent`
-
-- [x] **4.1** Temporal index — sorted timestamp index with binary search, insert/remove
-- [x] **4.2** Time-range queries — range_query, before, after, latest, earliest
-- [x] **4.3** Session DAG — parent/child linking, chain walking, time-range overlap queries
-- [x] **4.4** Temporal re-ranking — query hint enum (Latest/Earliest/Around/Between/None) with boost scoring
-- [x] **4.5** Temporal entity tracking — EntityTimeline with state change history + point-in-time reconstruction
-- [x] **4.6** Tests — comprehensive tests for all features
-
-**Research:** MemX temporal gaps (≤43.6% Hit@5), MemoryArena multi-session tasks (2026)
-
----
-
-## Track 5: Memory Security & Provenance
-**Status:** 🟢 Phase 1 Complete
-**Priority:** Medium-High
-**Crate:** `clawhdf5-agent`
-
-- [x] **5.1** Source attribution — MemoryProvenance with source, creator, session, FNV-1a content hash
-- [x] **5.2** Write anomaly detection — rate limiting, 15 injection patterns, source distribution analysis
-- [x] **5.3** Source isolation — per-MemorySource sub-stores preventing cross-contamination
-- [x] **5.4** Memory integrity verification — content hash comparison via verify_integrity()
-- [x] **5.5** Poisoning resistance — pattern detection for prompt injection attempts
-- [x] **5.6** Tests — comprehensive tests including adversarial patterns
-
-**Research:** MemoryGraft (2025), SSGM Framework (2026)
-
----
-
-## Track 6: Multi-Modal Memory
-**Status:** 🟢 Phase 1 Complete
-**Priority:** Medium
-**Crate:** `clawhdf5-agent`
-
-- [x] **6.1** Image embedding storage — ModalEmbedding with model provenance (CLIP, SigLIP, etc.)
-- [x] **6.2** Audio fingerprints — Audio modality with embedding storage
-- [x] **6.3** Multi-modal search — search_by_modality (filtered) + search_cross_modal (all embeddings)
-- [x] **6.4** Observation records — raw perception vs interpretation with confidence scoring
-- [x] **6.5** Media reference storage — MediaRef with Path/Url/Inline, MIME types, FNV-1a checksums
-- [x] **6.6** Tests — 35 comprehensive tests
-
-**Research:** Neuro-Symbolic Memory (2026), RAGdb multi-modal RAG (2025)
-
----
-
-## Track 7: OpenClaw Integration
-**Status:** 🟢 Complete
-**Priority:** Critical (for adoption)
-**Crates:** `clawhdf5-agent`, `clawhdf5-napi`
-
-- [x] **7.1** Memory backend trait — MemoryBackend with search/get/write/ingest/export/stats
-- [x] **7.2** Hybrid retrieval pipeline — ClawhdfBackend wires RRF → reranker → confidence rejection
-- [x] **7.3** Markdown import/export — MarkdownParser + MarkdownExporter with line tracking + metadata
-- [x] **7.4** memory_search tool — backed by full hybrid retrieval pipeline
-- [x] **7.5** memory_get tool — get() with path + line range support
-- [x] **7.6** Compaction integration — run_compaction() (decay + compact + WAL flush), run_consolidation() (hippocampal engine), tick_session(), flush_wal()
-- [x] **7.7** Config surface — `memory.backend = "clawhdf5"` schema documented in docs/openclaw-config.md
-- [x] **7.8** Documentation + migration guide — docs/migration-guide.md, docs/openclaw-integration.md (architecture, full API reference, code patterns)
-
-**Node.js bridge:** `clawhdf5-napi` (napi-rs) → `@redclaw/clawhdf5` npm package with full TypeScript types.
-
----
-
-## Track 8: Benchmarking & Validation
-**Status:** 🟢 Complete
-**Priority:** High
-**Crates:** `clawhdf5-agent`, `clawhdf5-bench`
-
-- [x] **8.1** MemoryArena benchmark — 35 queries, 50 sessions, Hit@10=91.4%, MRR=0.547
-- [x] **8.2** LongMemEval benchmark — 500 queries, session Hit@1=100%, turn Hit@5=84.4% (beats MemX 51.6%), MRR=0.660
-- [x] **8.3** Latency benchmarks — vector search at 1K/10K/100K, hybrid/RRF, graph traversal, consolidation, temporal
-- [x] **8.4** Memory footprint — 1.7 KB/record uncompressed, 282 B compressed (6.2x ratio), 100K+ rec/s ingestion
-- [x] **8.5** Consolidation efficiency — 8.8x search speedup, 90% noise eviction, zero quality loss
-- [x] **8.6** Cross-platform benchmarks — x86 measured, ARM estimated, cross_platform.sh script
-- [x] **8.7** Published results in BENCHMARKS.md with ephemeral tier Redis comparison (70-140x faster)
-
----
-
-## Implementation Order
-
-**Phase 1:** ~~Tracks 1, 2, 3 — core memory intelligence~~ 🟢 Complete
-**Phase 2:** ~~Track 4 (temporal) + Track 5 (security)~~ 🟢 Complete
-**Phase 3:** ~~Track 6 (multi-modal) + Track 7 (OpenClaw integration)~~ 🟢 Complete
-**Phase 4:** ~~Track 8 (benchmarking + validation)~~ 🟢 Complete
-
-All 8 tracks delivered. 1,650+ tests passing, zero clippy warnings.
-
----
-
-## What's Next
-
-Verified against current repo state on 2026-08-05 (see also `docs/superpowers/plans/` for the filter-codec/format-write/MPI-IO work, now shipped):
-
-- [ ] TypeScript bridge not wired into CI — `packages/clawhdf5-node/` already has a complete, working napi-rs package (package.json, tsconfig, hand-written TS wrapper matching all 21 `#[napi]` items, Jest test suite, README); it isn't published to npm and has no committed lockfile
-- [ ] Publish crates to crates.io — no `publish` config anywhere in the workspace yet
-- [ ] Python wheel distribution via maturin — `crates/clawhdf5-py/pyproject.toml` exists (maturin-buildable locally) but wheels aren't published anywhere
-- [ ] `chunked_read.rs`/`data_read.rs` full bounds-check audit + scheduled fuzz campaigns (the new `fuzz_dataset_read` target covers the two files' main entry points; a full manual audit of every indexing site is still open) — see Tier 4 below
-- [ ] WAL per-entry checksum landed as CRC32 (see below); a stronger per-entry format (explicit length prefix, avoiding the read-then-verify restructuring) could still be revisited if profiling shows it matters
-- [ ] HNSW build parallelism is still narrow (only `prune_connections`); the correctness-sensitive outer insert loop needs its own dedicated design pass before parallelizing
-
-### Recently closed out (2026-08-05, Tier 3–4 hardening pass)
-
-- [x] Academic benchmark cross-validation — LongMemEval reproduced against MemX on tank (Ryzen 7 7800X3D): turn-level Hit@5 84.4% vs MemX's 51.6%; recall numbers are deterministic and reproduce exactly across machines. SIMD/Parallelism and Vector Search sections also re-run and dated. See [BENCHMARKS.md § Independent Validation: tank — LongMemEval & Vector Search](BENCHMARKS.md#independent-validation-tank--longmemeval--vector-search-ryzen-7-7800x3d-2026-08-05)
-- [x] Android JNI (`clawhdf5-android`): validate `embedding_len`/`query_embedding_len` against the handle's configured `embedding_dim` before constructing a slice from a raw pointer
-- [x] `clawhdf5-py`: bumped pyo3/numpy 0.28 → 0.29, clearing two RUSTSEC advisories
-- [x] WAL (`clawhdf5-agent`): length-prefix caps (`MAX_WAL_FIELD_LEN`) to reject a corrupted length claim before allocating, then a full per-entry CRC32 trailer (`WAL_VERSION` 2) so a bit-flip stops replay cleanly instead of loading corrupted data; old-format WAL files still read correctly and are migrated on next open
-- [x] `chunked_read.rs`/`data_read.rs`/`local_heap.rs` bounds-check audit: added `ensure_len` overflow guards, a recursion-depth guard against cyclic B-trees, and a fix for an unguarded compound-datatype byte-offset overrun. Added a new `fuzz_dataset_read` cargo-fuzz target exercising the contiguous/chunked/compact read paths — it found and we fixed 3 real crash bugs (integer-overflow panics) within the first few runs
-- [x] `clawhdf5-ann`: optional `parallel` feature (rayon) for HNSW's `prune_connections` neighbor-distance computation
-- [x] `[workspace.dependencies]` added for `tempfile`/`criterion`/`half`/`serde`, fixing a real version skew on `half` (2 vs 2.7)
-
-### Recently closed out (2026-08-05 hardening pass)
-
-- [x] CI/CD pipeline — `.gitea/workflows/ci.yml` now runs `scripts/ci-test.sh` (fmt, clippy, tests, no_std check) on push/PR to `main`
-- [x] Fixed no_std build breakage in `clawhdf5-format` (missing alloc imports, `AtomicU64` unsupported on thumbv7em, `f64::powi` requiring std/libm)
-- [x] Fixed version skew: `clawhdf5-py` (pyproject.toml) and `packages/clawhdf5-node` (package.json) were both behind the actual crate version
-
-### Recently closed out (2026-08-03 cleanup pass)
-
-- [x] Removed `clawhdf5-types` — it was an empty 1-line stub crate; shared type definitions already live in `clawhdf5-format`, so CLAUDE.md and the workspace manifest were corrected instead of filling it in
-- [x] Superblock v4 (page-buffer mode) read/write — the only unimplemented task from `docs/superpowers/plans/2026-06-29-format-write-extensions.md`; now done (`Superblock::parse_v4`/`serialize`, `FileWriter::with_page_size`)
-- [x] Reconciled the three `docs/superpowers/plans/*.md` docs against actual shipped code — they were pre-work plans for `d6c4d4f` (2026-06-30), committed to git late; checkboxes now reflect reality
-
----
-
-_Last updated: 2026-08-05_
+The old track-by-track tracker this file used to be (agent-memory
+Tracks 1–8, mid-2026) is in git history (`git log -- ROADMAP.md`).

benchmarks/cross_platform.sh

@@ -29,7 +29,8 @@
 #     - Use wasm-pack with a custom bench harness
 #     - Replace std::time::Instant with web_sys::Performance::now()
 #     - Replace TempDir/HDF5 I/O with an in-memory backend (separate effort)
-#   See ROADMAP.md §WASM for the full scope.
+#   Browser reads are tested (not benchmarked) by
+#   examples/wasm-viewer/test/run.sh; see examples/wasm-viewer/README.md.
 
 set -euo pipefail
 

conformance/.gitignore

@@ -0,0 +1,3 @@
+/.cache/
+# pin the probe's dependencies (the workspace lock is not committed)
+!/probe/Cargo.lock

conformance/README.md

@@ -0,0 +1,70 @@
+# Conformance sweep
+
+Reads every HDF5 file of eight public corpora with clawhdf5 and with
+h5py/libhdf5, compares the two readings object by object, and writes
+[`CONFORMANCE.md`](../CONFORMANCE.md).
+
+```sh
+CLAWHDF5_PYTHON=/path/to/venv/bin/python conformance/run.sh      # ~30 s once the corpus is cached
+conformance/run.sh --no-fetch                                    # use the cached corpus as is
+conformance/run.sh --update-baseline                             # after an intended change in results
+```
+
+Latest result (tank, 2026-09-28 04:29 UTC, `conformance/run.sh --no-fetch
+--update-baseline`): 602 of 697 files ok, 1 our-error, 0 mismatch, 2
+ref-bug, 92 h5py-cannot-read, and no panic, hang, crash or out-of-memory.
+The our-error file is `bad_nbit_parms_walk.h5`, which flips between ref-bug
+and our-error from run to run (see `docs/known-issues.md`). The report with every file is
+[`CONFORMANCE.md`](../CONFORMANCE.md).
+
+## Classes
+
+`compare.py` puts each file in one class:
+
+| class | meaning |
+|---|---|
+| **ok** | clawhdf5 and h5py read the same objects with the same values |
+| **our-error** | h5py reads something clawhdf5 refuses |
+| **mismatch** | both read it, with different values or structure |
+| **h5py-cannot-read** | h5py (libhdf5) cannot read the file; not compared |
+| **ref-bug** | h5py reads an object clawhdf5 refuses, but only through a libhdf5 over-read: `ref_bugs.py` re-reads it in six processes with different heaps (import order, `MALLOC_PERTURB_`) and its values change. The file is ref-bug only while that is confirmed in the same run; if the values become stable it counts as our-error again |
+| **panic / hang / crash / oom** | a clawhdf5 failure under the timeout and address-space limit; the gate fails on any |
+
+Where h5py itself returns wrong values through a known h5py bug (the
+big-endian variable-length bug: elements returned with the file's bytes
+under a little-endian dtype), `ref.py` checks that the installed h5py has
+the bug, corrects the values before hashing and marks them `ref_fix`, so
+those objects are still compared. The evidence for the three remaining
+non-ok files (ref-bug or, for one, our-error) is under "Conformance: the last non-ok files" in
+[`docs/known-issues.md`](../docs/known-issues.md).
+
+Needs Rust, `git`, `h5dump` (Debian/Ubuntu `hdf5-tools`), `libaec` (for the
+probe's `szip` feature; `libaec-dev`), and a Python with the packages in
+`requirements.txt`. The first run downloads about 450 MB of sparse checkouts.
+
+| file | role |
+|---|---|
+| `corpus.txt` | the corpora: git URL, pinned commit, swept root, sparse-checkout patterns |
+| `fetch-corpus.sh` | shallow, sparse, blob-filtered checkout of each pinned commit into `.cache/src/` (gitignored); no-op when already there |
+| `list_files.py` | which files are probed (HDF5/netCDF-4 extensions minus netCDF classic, plus the CVE reproducers) |
+| `probe/` | the clawhdf5 side: a standalone crate (outside the workspace, so `cargo test --workspace` never builds it) that walks a file with `clawhdf5-format` and prints canonical JSON |
+| `ref.py` | the h5py side: the same JSON from h5py (values corrected for a known h5py bug are marked `ref_fix`) |
+| `ref_bugs.py` | re-reads the objects h5py reads only through a libhdf5 bug in six differently-set-up processes; an object whose values change is confirmed as a libhdf5 over-read |
+| `test_ref.py` | tests of `ref.py`'s correction and `ref_bugs.py`'s confirmation (`python conformance/test_ref.py`) |
+| `run_one.sh` | runs both sides on one file (and `h5dump` on the CVE corpus) under a timeout and an address-space limit |
+| `compare.py` | classifies each file (ok / our-error / mismatch / h5py-cannot-read / ref-bug / panic / hang / crash / oom) and groups root causes |
+| `report.py` | writes `CONFORMANCE.md` |
+| `check.py` | the gate: fails on any panic/hang/crash/oom, on an ok count below `baseline.json`, or on a baseline-ok file that is no longer ok |
+| `baseline.json` | the ok files the gate holds the line on |
+| `requirements.txt` | pinned h5py / numpy / hdf5plugin / netCDF4 |
+
+Results for every file (both sides' JSON and stderr, `results.csv`,
+`results.json`, `summary.md`) are left in `.cache/results/`.
+
+The nightly job is `.gitea/workflows/conformance.yml`; it prints the report
+into the job log.
+
+The canonical value encoding both sides hash is documented at the top of
+`probe/src/main.rs`. Values are compared as libhdf5 presents them: a float
+with a non-IEEE bit layout (N-Bit) or an integer with a bit offset is compared
+as the converted number, not as raw file bytes.

conformance/baseline.json

@@ -0,0 +1,648 @@
+{
+ "comment": "conformance/run.sh fails if the ok count drops below `ok` or a file in `ok_files` stops being ok. Regenerate with `conformance/run.sh --update-baseline` after an intended change.",
+ "commit": "bf5a163dcf7fe28d651ada6545d8136ffeffc825",
+ "date": "2026-09-28 04:29 UTC",
+ "reference": "h5py 3.16.0 / HDF5 2.0.0",
+ "files": 697,
+ "ok": 602,
+ "counts": {
+  "h5py-cannot-read": 92,
+  "ok": 602,
+  "our-error": 1,
+  "ref-bug": 2
+ },
+ "per_corpus": {
+  "NCAS-CMS_pyfive": {
+   "ok": 33
+  },
+  "cve_hdf5": {
+   "h5py-cannot-read": 32,
+   "ok": 113,
+   "ref-bug": 2
+  },
+  "h5py_data": {
+   "ok": 4
+  },
+  "hdf5": {
+   "h5py-cannot-read": 60,
+   "ok": 405,
+   "our-error": 1
+  },
+  "netcdf-c": {
+   "ok": 20
+  },
+  "netcdf4-python": {
+   "ok": 18
+  },
+  "usnistgov_h5wasm": {
+   "ok": 5
+  },
+  "xarray-data": {
+   "ok": 4
+  }
+ },
+ "ok_files": [
+  "NCAS-CMS_pyfive/tests/compact.hdf5",
+  "NCAS-CMS_pyfive/tests/data/attr_datatypes.hdf5",
+  "NCAS-CMS_pyfive/tests/data/btreev2.hdf5",
+  "NCAS-CMS_pyfive/tests/data/chunked.hdf5",
+  "NCAS-CMS_pyfive/tests/data/cmip_bad_eg.nc",
+  "NCAS-CMS_pyfive/tests/data/compressed.hdf5",
+  "NCAS-CMS_pyfive/tests/data/compressed_v1.hdf5",
+  "NCAS-CMS_pyfive/tests/data/dataset_datatypes.hdf5",
+  "NCAS-CMS_pyfive/tests/data/dataset_multidim.hdf5",
+  "NCAS-CMS_pyfive/tests/data/dim_scales.hdf5",
+  "NCAS-CMS_pyfive/tests/data/earliest.hdf5",
+  "NCAS-CMS_pyfive/tests/data/enum_h5variable.hdf5",
+  "NCAS-CMS_pyfive/tests/data/enum_variable.hdf5",
+  "NCAS-CMS_pyfive/tests/data/enum_variable.nc",
+  "NCAS-CMS_pyfive/tests/data/enums_from_netcdf.nc",
+  "NCAS-CMS_pyfive/tests/data/fillvalue_earliest.hdf5",
+  "NCAS-CMS_pyfive/tests/data/fillvalue_latest.hdf5",
+  "NCAS-CMS_pyfive/tests/data/filter_pipeline_v2.hdf5",
+  "NCAS-CMS_pyfive/tests/data/fletcher32.hdf5",
+  "NCAS-CMS_pyfive/tests/data/fractal_heap_no_mci_rlat.nc",
+  "NCAS-CMS_pyfive/tests/data/groups.hdf5",
+  "NCAS-CMS_pyfive/tests/data/h5netcdf_test.hdf5",
+  "NCAS-CMS_pyfive/tests/data/issue23_A.nc",
+  "NCAS-CMS_pyfive/tests/data/issue23_A_contiguous.nc",
+  "NCAS-CMS_pyfive/tests/data/issue23_B.nc",
+  "NCAS-CMS_pyfive/tests/data/latest.hdf5",
+  "NCAS-CMS_pyfive/tests/data/netcdf4_classic.nc",
+  "NCAS-CMS_pyfive/tests/data/new_style_groups.hdf5",
+  "NCAS-CMS_pyfive/tests/data/noy_AERmonZ_UKESM1-0-LL_piControl_r1i1p1f2_gnz_200001-200012.nc",
+  "NCAS-CMS_pyfive/tests/data/references.hdf5",
+  "NCAS-CMS_pyfive/tests/data/resizable.hdf5",
+  "NCAS-CMS_pyfive/tests/opaque_datetime.hdf5",
+  "NCAS-CMS_pyfive/tests/opaque_fixed.hdf5",
+  "cve_hdf5/cvefiles/cve-2016-4330.h5",
+  "cve_hdf5/cvefiles/cve-2016-4331.h5",
+  "cve_hdf5/cvefiles/cve-2016-4332-mtime-new.h5",
+  "cve_hdf5/cvefiles/cve-2016-4332-mtime.h5",
+  "cve_hdf5/cvefiles/cve-2016-4333.h5",
+  "cve_hdf5/cvefiles/cve-2017-17505.h5",
+  "cve_hdf5/cvefiles/cve-2017-17506.h5",
+  "cve_hdf5/cvefiles/cve-2017-17507.h5",
+  "cve_hdf5/cvefiles/cve-2017-17508.h5",
+  "cve_hdf5/cvefiles/cve-2017-17509.h5",
+  "cve_hdf5/cvefiles/cve-2018-11202.h5",
+  "cve_hdf5/cvefiles/cve-2018-11203.h5",
+  "cve_hdf5/cvefiles/cve-2018-11204.h5",
+  "cve_hdf5/cvefiles/cve-2018-11205.h5",
+  "cve_hdf5/cvefiles/cve-2018-11206-new.h5",
+  "cve_hdf5/cvefiles/cve-2018-11206-old.h5",
+  "cve_hdf5/cvefiles/cve-2018-11207.h5",
+  "cve_hdf5/cvefiles/cve-2018-13867.h5",
+  "cve_hdf5/cvefiles/cve-2018-13868.h5",
+  "cve_hdf5/cvefiles/cve-2018-13869.h5",
+  "cve_hdf5/cvefiles/cve-2018-13870.h5",
+  "cve_hdf5/cvefiles/cve-2018-13871.h5",
+  "cve_hdf5/cvefiles/cve-2018-13872.h5",
+  "cve_hdf5/cvefiles/cve-2018-13873.h5",
+  "cve_hdf5/cvefiles/cve-2018-13875.h5",
+  "cve_hdf5/cvefiles/cve-2018-14031.h5",
+  "cve_hdf5/cvefiles/cve-2018-14033.h5",
+  "cve_hdf5/cvefiles/cve-2018-14034.h5",
+  "cve_hdf5/cvefiles/cve-2018-14035.h5",
+  "cve_hdf5/cvefiles/cve-2018-14460.h5",
+  "cve_hdf5/cvefiles/cve-2018-15671.h5",
+  "cve_hdf5/cvefiles/cve-2018-15672.h5",
+  "cve_hdf5/cvefiles/cve-2018-16438.h5",
+  "cve_hdf5/cvefiles/cve-2018-17233.h5",
+  "cve_hdf5/cvefiles/cve-2018-17234.h5",
+  "cve_hdf5/cvefiles/cve-2018-17237.h5",
+  "cve_hdf5/cvefiles/cve-2018-17432.h5",
+  "cve_hdf5/cvefiles/cve-2018-17434.h5",
+  "cve_hdf5/cvefiles/cve-2018-17435.h5",
+  "cve_hdf5/cvefiles/cve-2018-17437.h5",
+  "cve_hdf5/cvefiles/cve-2018-17438",
+  "cve_hdf5/cvefiles/cve-2018-17439",
+  "cve_hdf5/cvefiles/cve-2019-8396.h5",
+  "cve_hdf5/cvefiles/cve-2019-8397.h5",
+  "cve_hdf5/cvefiles/cve-2019-8398.h5",
+  "cve_hdf5/cvefiles/cve-2019-9151.h5",
+  "cve_hdf5/cvefiles/cve-2019-9152.h5",
+  "cve_hdf5/cvefiles/cve-2020-10811.h5",
+  "cve_hdf5/cvefiles/cve-2020-18232.h5",
+  "cve_hdf5/cvefiles/cve-2020-18494.h5",
+  "cve_hdf5/cvefiles/cve-2021-36977.h5",
+  "cve_hdf5/cvefiles/cve-2021-37501.h5",
+  "cve_hdf5/cvefiles/cve-2021-45829.h5",
+  "cve_hdf5/cvefiles/cve-2021-45833.h5",
+  "cve_hdf5/cvefiles/cve-2021-46243.h5",
+  "cve_hdf5/cvefiles/cve-2021-46244.h5",
+  "cve_hdf5/cvefiles/cve-2024-29157.h5",
+  "cve_hdf5/cvefiles/cve-2024-29158.h5",
+  "cve_hdf5/cvefiles/cve-2024-29159.h5",
+  "cve_hdf5/cvefiles/cve-2024-29160.h5",
+  "cve_hdf5/cvefiles/cve-2024-29161.h5",
+  "cve_hdf5/cvefiles/cve-2024-29162.h5",
+  "cve_hdf5/cvefiles/cve-2024-29163.h5",
+  "cve_hdf5/cvefiles/cve-2024-29164.h5",
+  "cve_hdf5/cvefiles/cve-2024-29165.h5",
+  "cve_hdf5/cvefiles/cve-2024-29166.h5",
+  "cve_hdf5/cvefiles/cve-2024-32605.h5",
+  "cve_hdf5/cvefiles/cve-2024-32606.h5",
+  "cve_hdf5/cvefiles/cve-2024-32607-1.h5",
+  "cve_hdf5/cvefiles/cve-2024-32607-2.h5",
+  "cve_hdf5/cvefiles/cve-2024-32608.h5",
+  "cve_hdf5/cvefiles/cve-2024-32610.h5",
+  "cve_hdf5/cvefiles/cve-2024-32611.h5",
+  "cve_hdf5/cvefiles/cve-2024-32612.h5",
+  "cve_hdf5/cvefiles/cve-2024-32613.h5",
+  "cve_hdf5/cvefiles/cve-2024-32614.h5",
+  "cve_hdf5/cvefiles/cve-2024-32615.h5",
+  "cve_hdf5/cvefiles/cve-2024-32616.h5",
+  "cve_hdf5/cvefiles/cve-2024-32617.h5",
+  "cve_hdf5/cvefiles/cve-2024-32618.h5",
+  "cve_hdf5/cvefiles/cve-2024-32619.h5",
+  "cve_hdf5/cvefiles/cve-2024-32620.h5",
+  "cve_hdf5/cvefiles/cve-2024-32621.h5",
+  "cve_hdf5/cvefiles/cve-2024-32622.h5",
+  "cve_hdf5/cvefiles/cve-2024-32623.h5",
+  "cve_hdf5/cvefiles/cve-2024-32624.h5",
+  "cve_hdf5/cvefiles/cve-2024-33873.h5",
+  "cve_hdf5/cvefiles/cve-2024-33874.h5",
+  "cve_hdf5/cvefiles/cve-2024-33875.h5",
+  "cve_hdf5/cvefiles/cve-2024-33876.h5",
+  "cve_hdf5/cvefiles/cve-2024-33877.h5",
+  "cve_hdf5/cvefiles/cve-2025-2309.h5",
+  "cve_hdf5/cvefiles/cve-2025-2310.h5",
+  "cve_hdf5/cvefiles/cve-2025-2924.h5",
+  "cve_hdf5/cvefiles/cve-2025-2925.h5",
+  "cve_hdf5/cvefiles/cve-2025-44905.h5",
+  "cve_hdf5/cvefiles/cve-2025-6269-1.h5",
+  "cve_hdf5/cvefiles/cve-2025-6269-2.h5",
+  "cve_hdf5/cvefiles/cve-2025-6269-3.h5",
+  "cve_hdf5/cvefiles/cve-2025-6269-4.h5",
+  "cve_hdf5/cvefiles/cve-2025-6516.h5",
+  "cve_hdf5/cvefiles/cve-2025-6857.h5",
+  "cve_hdf5/cvefiles/cve-2025-7067.h5",
+  "cve_hdf5/cvefiles/cve-2026-26200.h5",
+  "cve_hdf5/cvefiles/cve-2026-34734.h5",
+  "cve_hdf5/cvefiles/cve-2026-92627.h5",
+  "cve_hdf5/cvefiles/unknown-1.h5",
+  "cve_hdf5/fuzzerfiles/gh-4431-poc-03.h5",
+  "cve_hdf5/fuzzerfiles/gh-4432-poc-05.h5",
+  "cve_hdf5/fuzzerfiles/gh-4433-poc-08.h5",
+  "cve_hdf5/fuzzerfiles/gh-4435-poc-10.h5",
+  "cve_hdf5/fuzzerfiles/gh_2649_flawed.h5",
+  "cve_hdf5/fuzzerfiles/gh_2649_plain_model.h5",
+  "h5py_data/compound-dtype-complex.h5",
+  "h5py_data/vlen_string_dset.h5",
+  "h5py_data/vlen_string_dset_utc.h5",
+  "h5py_data/vlen_string_s390x.h5",
+  "hdf5/HDF5Examples/C/H5FLT/tfiles/h5ex_d_bitgroom.h5",
+  "hdf5/HDF5Examples/C/H5FLT/tfiles/h5ex_d_blosc.h5",
+  "hdf5/HDF5Examples/C/H5FLT/tfiles/h5ex_d_blosc2.h5",
+  "hdf5/HDF5Examples/C/H5FLT/tfiles/h5ex_d_bshuf.h5",
+  "hdf5/HDF5Examples/C/H5FLT/tfiles/h5ex_d_bzip2.h5",
+  "hdf5/HDF5Examples/C/H5FLT/tfiles/h5ex_d_granularbr.h5",
+  "hdf5/HDF5Examples/C/H5FLT/tfiles/h5ex_d_jpeg.h5",
+  "hdf5/HDF5Examples/C/H5FLT/tfiles/h5ex_d_lz4.h5",
+  "hdf5/HDF5Examples/C/H5FLT/tfiles/h5ex_d_lzf.h5",
+  "hdf5/HDF5Examples/C/H5FLT/tfiles/h5ex_d_zfp.h5",
+  "hdf5/HDF5Examples/C/H5FLT/tfiles/h5ex_d_zstd.h5",
+  "hdf5/HDF5Examples/C/H5G/16/h5ex_g_iterate.h5",
+  "hdf5/HDF5Examples/C/H5G/16/h5ex_g_traverse.h5",
+  "hdf5/HDF5Examples/C/H5G/h5ex_g_iterate.h5",
+  "hdf5/HDF5Examples/C/H5G/h5ex_g_traverse.h5",
+  "hdf5/HDF5Examples/C/H5G/h5ex_g_visit.h5",
+  "hdf5/HDF5Examples/FORTRAN/H5G/h5ex_g_iterate.h5",
+  "hdf5/HDF5Examples/FORTRAN/H5G/h5ex_g_traverse.h5",
+  "hdf5/HDF5Examples/FORTRAN/H5G/h5ex_g_visit.h5",
+  "hdf5/HDF5Examples/JAVA/H5G/h5ex_g_iterate.h5",
+  "hdf5/HDF5Examples/JAVA/H5G/h5ex_g_visit.h5",
+  "hdf5/HDF5Examples/JAVA/compat/H5G/110/h5ex_g_iterate.h5",
+  "hdf5/HDF5Examples/JAVA/compat/H5G/110/h5ex_g_visit.h5",
+  "hdf5/HDF5Examples/JAVA/compat/H5G/h5ex_g_iterate.h5",
+  "hdf5/HDF5Examples/JAVA/compat/H5G/h5ex_g_visit.h5",
+  "hdf5/c++/test/th5s.h5",
+  "hdf5/hl/test/testfiles/test_ds_be.h5",
+  "hdf5/hl/test/testfiles/test_ds_be_new_ref-32bit.h5",
+  "hdf5/hl/test/testfiles/test_ds_be_new_ref.h5",
+  "hdf5/hl/test/testfiles/test_ds_le.h5",
+  "hdf5/hl/test/testfiles/test_ds_le_new_ref.h5",
+  "hdf5/hl/test/testfiles/test_ld.h5",
+  "hdf5/hl/test/testfiles/test_table_be.h5",
+  "hdf5/hl/test/testfiles/test_table_cray.h5",
+  "hdf5/hl/test/testfiles/test_table_le.h5",
+  "hdf5/test/testfiles/aggr.h5",
+  "hdf5/test/testfiles/bad_chunk_ndims.h5",
+  "hdf5/test/testfiles/bad_compound.h5",
+  "hdf5/test/testfiles/bad_offset.h5",
+  "hdf5/test/testfiles/be_data.h5",
+  "hdf5/test/testfiles/be_extlink1.h5",
+  "hdf5/test/testfiles/be_extlink2.h5",
+  "hdf5/test/testfiles/btree_idx_1_6.h5",
+  "hdf5/test/testfiles/btree_idx_1_8.h5",
+  "hdf5/test/testfiles/charsets.h5",
+  "hdf5/test/testfiles/corrupt_stab_msg.h5",
+  "hdf5/test/testfiles/deflate.h5",
+  "hdf5/test/testfiles/file_image_core_test.h5",
+  "hdf5/test/testfiles/filespace_1_6.h5",
+  "hdf5/test/testfiles/filespace_1_8.h5",
+  "hdf5/test/testfiles/fill18.h5",
+  "hdf5/test/testfiles/fill_old.h5",
+  "hdf5/test/testfiles/filter_error.h5",
+  "hdf5/test/testfiles/fsm_aggr_nopersist.h5",
+  "hdf5/test/testfiles/fsm_aggr_persist.h5",
+  "hdf5/test/testfiles/group_old.h5",
+  "hdf5/test/testfiles/h5fc_ext1_f.h5",
+  "hdf5/test/testfiles/h5fc_ext1_i.h5",
+  "hdf5/test/testfiles/h5fc_ext2_if.h5",
+  "hdf5/test/testfiles/h5fc_ext2_sf.h5",
+  "hdf5/test/testfiles/h5fc_ext3_isf.h5",
+  "hdf5/test/testfiles/h5fc_ext_none.h5",
+  "hdf5/test/testfiles/le_data.h5",
+  "hdf5/test/testfiles/le_extlink1.h5",
+  "hdf5/test/testfiles/le_extlink2.h5",
+  "hdf5/test/testfiles/memleak_H5O_dtype_decode_helper_H5Odtype.h5",
+  "hdf5/test/testfiles/mergemsg.h5",
+  "hdf5/test/testfiles/noencoder.h5",
+  "hdf5/test/testfiles/none.h5",
+  "hdf5/test/testfiles/paged_nopersist.h5",
+  "hdf5/test/testfiles/paged_persist.h5",
+  "hdf5/test/testfiles/specmetaread.h5",
+  "hdf5/test/testfiles/tarrold.h5",
+  "hdf5/test/testfiles/tbad_msg_count.h5",
+  "hdf5/test/testfiles/tbogus.h5",
+  "hdf5/test/testfiles/test_filters_be.h5",
+  "hdf5/test/testfiles/test_filters_le.h5",
+  "hdf5/test/testfiles/th5s.h5",
+  "hdf5/test/testfiles/tlayouto.h5",
+  "hdf5/test/testfiles/tmisc38a.h5",
+  "hdf5/test/testfiles/tmisc38b.h5",
+  "hdf5/test/testfiles/tmtimen.h5",
+  "hdf5/test/testfiles/tmtimeo.h5",
+  "hdf5/test/testfiles/tnullspace.h5",
+  "hdf5/test/testfiles/tsizeslheap.h5",
+  "hdf5/tools/test/testfiles/bigendian/tall.h5",
+  "hdf5/tools/test/testfiles/bigendian/tdset2.h5",
+  "hdf5/tools/test/testfiles/binfp64.h5",
+  "hdf5/tools/test/testfiles/binin16.h5",
+  "hdf5/tools/test/testfiles/binin32.h5",
+  "hdf5/tools/test/testfiles/binin8.h5",
+  "hdf5/tools/test/testfiles/binin8w.h5",
+  "hdf5/tools/test/testfiles/binuin16.h5",
+  "hdf5/tools/test/testfiles/binuin32.h5",
+  "hdf5/tools/test/testfiles/bounds_latest_latest.h5",
+  "hdf5/tools/test/testfiles/charsets.h5",
+  "hdf5/tools/test/testfiles/compounds_array_vlen1.h5",
+  "hdf5/tools/test/testfiles/compounds_array_vlen2.h5",
+  "hdf5/tools/test/testfiles/err_attr_dspace.h5",
+  "hdf5/tools/test/testfiles/file_space.h5",
+  "hdf5/tools/test/testfiles/filter_fail.h5",
+  "hdf5/tools/test/testfiles/h5clear_fsm_persist_equal.h5",
+  "hdf5/tools/test/testfiles/h5clear_fsm_persist_less.h5",
+  "hdf5/tools/test/testfiles/h5clear_fsm_persist_noclose.h5",
+  "hdf5/tools/test/testfiles/h5clear_fsm_persist_user_equal.h5",
+  "hdf5/tools/test/testfiles/h5clear_fsm_persist_user_less.h5",
+  "hdf5/tools/test/testfiles/h5clear_mdc_image.h5",
+  "hdf5/tools/test/testfiles/h5clear_sec2_v0.h5",
+  "hdf5/tools/test/testfiles/h5clear_sec2_v2.h5",
+  "hdf5/tools/test/testfiles/h5copy_extlinks_src.h5",
+  "hdf5/tools/test/testfiles/h5copy_extlinks_trg.h5",
+  "hdf5/tools/test/testfiles/h5copy_ref.h5",
+  "hdf5/tools/test/testfiles/h5copytst.h5",
+  "hdf5/tools/test/testfiles/h5copytst_new.h5",
+  "hdf5/tools/test/testfiles/h5diff_attr1.h5",
+  "hdf5/tools/test/testfiles/h5diff_attr2.h5",
+  "hdf5/tools/test/testfiles/h5diff_attr3.h5",
+  "hdf5/tools/test/testfiles/h5diff_attr_v_level1.h5",
+  "hdf5/tools/test/testfiles/h5diff_attr_v_level2.h5",
+  "hdf5/tools/test/testfiles/h5diff_basic1.h5",
+  "hdf5/tools/test/testfiles/h5diff_basic2.h5",
+  "hdf5/tools/test/testfiles/h5diff_comp_vl_strs.h5",
+  "hdf5/tools/test/testfiles/h5diff_danglelinks1.h5",
+  "hdf5/tools/test/testfiles/h5diff_danglelinks2.h5",
+  "hdf5/tools/test/testfiles/h5diff_dset1.h5",
+  "hdf5/tools/test/testfiles/h5diff_dset2.h5",
+  "hdf5/tools/test/testfiles/h5diff_dset3.h5",
+  "hdf5/tools/test/testfiles/h5diff_dset_zero_dim_size1.h5",
+  "hdf5/tools/test/testfiles/h5diff_dset_zero_dim_size2.h5",
+  "hdf5/tools/test/testfiles/h5diff_dtypes.h5",
+  "hdf5/tools/test/testfiles/h5diff_empty.h5",
+  "hdf5/tools/test/testfiles/h5diff_enum_invalid_values.h5",
+  "hdf5/tools/test/testfiles/h5diff_eps1.h5",
+  "hdf5/tools/test/testfiles/h5diff_eps2.h5",
+  "hdf5/tools/test/testfiles/h5diff_exclude1-1.h5",
+  "hdf5/tools/test/testfiles/h5diff_exclude1-2.h5",
+  "hdf5/tools/test/testfiles/h5diff_exclude2-1.h5",
+  "hdf5/tools/test/testfiles/h5diff_exclude2-2.h5",
+  "hdf5/tools/test/testfiles/h5diff_exclude3-1.h5",
+  "hdf5/tools/test/testfiles/h5diff_exclude3-2.h5",
+  "hdf5/tools/test/testfiles/h5diff_ext2softlink_src.h5",
+  "hdf5/tools/test/testfiles/h5diff_ext2softlink_trg.h5",
+  "hdf5/tools/test/testfiles/h5diff_extlink_src.h5",
+  "hdf5/tools/test/testfiles/h5diff_extlink_trg.h5",
+  "hdf5/tools/test/testfiles/h5diff_grp_recurse1.h5",
+  "hdf5/tools/test/testfiles/h5diff_grp_recurse2.h5",
+  "hdf5/tools/test/testfiles/h5diff_grp_recurse_ext1.h5",
+  "hdf5/tools/test/testfiles/h5diff_grp_recurse_ext2-1.h5",
+  "hdf5/tools/test/testfiles/h5diff_grp_recurse_ext2-2.h5",
+  "hdf5/tools/test/testfiles/h5diff_grp_recurse_ext2-3.h5",
+  "hdf5/tools/test/testfiles/h5diff_hyper1.h5",
+  "hdf5/tools/test/testfiles/h5diff_hyper2.h5",
+  "hdf5/tools/test/testfiles/h5diff_linked_softlink.h5",
+  "hdf5/tools/test/testfiles/h5diff_links.h5",
+  "hdf5/tools/test/testfiles/h5diff_onion_dset_1d.h5",
+  "hdf5/tools/test/testfiles/h5diff_onion_dset_ext.h5",
+  "hdf5/tools/test/testfiles/h5diff_onion_objs.h5",
+  "hdf5/tools/test/testfiles/h5diff_softlinks.h5",
+  "hdf5/tools/test/testfiles/h5diff_strings1.h5",
+  "hdf5/tools/test/testfiles/h5diff_strings2.h5",
+  "hdf5/tools/test/testfiles/h5diff_types.h5",
+  "hdf5/tools/test/testfiles/h5fc_edge_v3.h5",
+  "hdf5/tools/test/testfiles/h5fc_err_level.h5",
+  "hdf5/tools/test/testfiles/h5fc_ext1_f.h5",
+  "hdf5/tools/test/testfiles/h5fc_ext1_i.h5",
+  "hdf5/tools/test/testfiles/h5fc_ext1_s.h5",
+  "hdf5/tools/test/testfiles/h5fc_ext2_if.h5",
+  "hdf5/tools/test/testfiles/h5fc_ext2_is.h5",
+  "hdf5/tools/test/testfiles/h5fc_ext2_sf.h5",
+  "hdf5/tools/test/testfiles/h5fc_ext3_isf.h5",
+  "hdf5/tools/test/testfiles/h5fc_ext_none.h5",
+  "hdf5/tools/test/testfiles/h5fc_non_v3.h5",
+  "hdf5/tools/test/testfiles/h5repack_CVE-2018-14460.h5",
+  "hdf5/tools/test/testfiles/h5repack_CVE-2018-17432.h5",
+  "hdf5/tools/test/testfiles/h5repack_aggr.h5",
+  "hdf5/tools/test/testfiles/h5repack_attr.h5",
+  "hdf5/tools/test/testfiles/h5repack_attr_refs.h5",
+  "hdf5/tools/test/testfiles/h5repack_deflate.h5",
+  "hdf5/tools/test/testfiles/h5repack_early.h5",
+  "hdf5/tools/test/testfiles/h5repack_ext.h5",
+  "hdf5/tools/test/testfiles/h5repack_f32le.h5",
+  "hdf5/tools/test/testfiles/h5repack_f32le_ex.h5",
+  "hdf5/tools/test/testfiles/h5repack_fill.h5",
+  "hdf5/tools/test/testfiles/h5repack_filters.h5",
+  "hdf5/tools/test/testfiles/h5repack_fletcher.h5",
+  "hdf5/tools/test/testfiles/h5repack_fsm_aggr_nopersist.h5",
+  "hdf5/tools/test/testfiles/h5repack_fsm_aggr_persist.h5",
+  "hdf5/tools/test/testfiles/h5repack_hlink.h5",
+  "hdf5/tools/test/testfiles/h5repack_int32le_1d.h5",
+  "hdf5/tools/test/testfiles/h5repack_int32le_1d_ex.h5",
+  "hdf5/tools/test/testfiles/h5repack_int32le_2d.h5",
+  "hdf5/tools/test/testfiles/h5repack_int32le_2d_ex.h5",
+  "hdf5/tools/test/testfiles/h5repack_int32le_3d.h5",
+  "hdf5/tools/test/testfiles/h5repack_int32le_3d_ex.h5",
+  "hdf5/tools/test/testfiles/h5repack_layout.UD.h5",
+  "hdf5/tools/test/testfiles/h5repack_layout.h5",
+  "hdf5/tools/test/testfiles/h5repack_layout2.h5",
+  "hdf5/tools/test/testfiles/h5repack_layout3.h5",
+  "hdf5/tools/test/testfiles/h5repack_layouto.h5",
+  "hdf5/tools/test/testfiles/h5repack_named_dtypes.h5",
+  "hdf5/tools/test/testfiles/h5repack_nbit.h5",
+  "hdf5/tools/test/testfiles/h5repack_nested_8bit_enum.h5",
+  "hdf5/tools/test/testfiles/h5repack_nested_8bit_enum_deflated.h5",
+  "hdf5/tools/test/testfiles/h5repack_none.h5",
+  "hdf5/tools/test/testfiles/h5repack_objs.h5",
+  "hdf5/tools/test/testfiles/h5repack_paged_nopersist.h5",
+  "hdf5/tools/test/testfiles/h5repack_paged_persist.h5",
+  "hdf5/tools/test/testfiles/h5repack_refs.h5",
+  "hdf5/tools/test/testfiles/h5repack_shuffle.h5",
+  "hdf5/tools/test/testfiles/h5repack_soffset.h5",
+  "hdf5/tools/test/testfiles/h5repack_szip.h5",
+  "hdf5/tools/test/testfiles/h5repack_uint8be.h5",
+  "hdf5/tools/test/testfiles/h5repack_uint8be_ex.h5",
+  "hdf5/tools/test/testfiles/h5stat_err_old_fill.h5",
+  "hdf5/tools/test/testfiles/h5stat_err_old_layout.h5",
+  "hdf5/tools/test/testfiles/h5stat_err_refcount.h5",
+  "hdf5/tools/test/testfiles/h5stat_filters.h5",
+  "hdf5/tools/test/testfiles/h5stat_idx.h5",
+  "hdf5/tools/test/testfiles/h5stat_newgrat.h5",
+  "hdf5/tools/test/testfiles/h5stat_threshold.h5",
+  "hdf5/tools/test/testfiles/h5stat_tsohm.h5",
+  "hdf5/tools/test/testfiles/mod_h5clear_mdc_image.h5",
+  "hdf5/tools/test/testfiles/non_comparables1.h5",
+  "hdf5/tools/test/testfiles/non_comparables2.h5",
+  "hdf5/tools/test/testfiles/old_h5fc_ext1_f.h5",
+  "hdf5/tools/test/testfiles/old_h5fc_ext1_i.h5",
+  "hdf5/tools/test/testfiles/old_h5fc_ext1_s.h5",
+  "hdf5/tools/test/testfiles/old_h5fc_ext2_if.h5",
+  "hdf5/tools/test/testfiles/old_h5fc_ext2_is.h5",
+  "hdf5/tools/test/testfiles/old_h5fc_ext2_sf.h5",
+  "hdf5/tools/test/testfiles/old_h5fc_ext3_isf.h5",
+  "hdf5/tools/test/testfiles/old_h5fc_ext_none.h5",
+  "hdf5/tools/test/testfiles/packedbits.h5",
+  "hdf5/tools/test/testfiles/t128bit_float.h5",
+  "hdf5/tools/test/testfiles/tCVE-2021-37501_attr_decode.h5",
+  "hdf5/tools/test/testfiles/tCVE_2018_11206_fill_new.h5",
+  "hdf5/tools/test/testfiles/tCVE_2018_11206_fill_old.h5",
+  "hdf5/tools/test/testfiles/taindices.h5",
+  "hdf5/tools/test/testfiles/tall.h5",
+  "hdf5/tools/test/testfiles/tarray1.h5",
+  "hdf5/tools/test/testfiles/tarray1_big.h5",
+  "hdf5/tools/test/testfiles/tarray2.h5",
+  "hdf5/tools/test/testfiles/tarray3.h5",
+  "hdf5/tools/test/testfiles/tarray4.h5",
+  "hdf5/tools/test/testfiles/tarray5.h5",
+  "hdf5/tools/test/testfiles/tarray8.h5",
+  "hdf5/tools/test/testfiles/tattr.h5",
+  "hdf5/tools/test/testfiles/tattr2.h5",
+  "hdf5/tools/test/testfiles/tattr4_be.h5",
+  "hdf5/tools/test/testfiles/tattrintsize.h5",
+  "hdf5/tools/test/testfiles/tattrreg.h5",
+  "hdf5/tools/test/testfiles/tbfloat16.h5",
+  "hdf5/tools/test/testfiles/tbfloat16_be.h5",
+  "hdf5/tools/test/testfiles/tbigdims.h5",
+  "hdf5/tools/test/testfiles/tbinary.h5",
+  "hdf5/tools/test/testfiles/tbitnopaque.h5",
+  "hdf5/tools/test/testfiles/tchar.h5",
+  "hdf5/tools/test/testfiles/tcmpdattrintsize.h5",
+  "hdf5/tools/test/testfiles/tcmpdintarray.h5",
+  "hdf5/tools/test/testfiles/tcmpdints.h5",
+  "hdf5/tools/test/testfiles/tcmpdintsize.h5",
+  "hdf5/tools/test/testfiles/tcomplex.h5",
+  "hdf5/tools/test/testfiles/tcomplex_be.h5",
+  "hdf5/tools/test/testfiles/tcompound.h5",
+  "hdf5/tools/test/testfiles/tcompound_complex.h5",
+  "hdf5/tools/test/testfiles/tcompound_complex2.h5",
+  "hdf5/tools/test/testfiles/tdatareg.h5",
+  "hdf5/tools/test/testfiles/tdset.h5",
+  "hdf5/tools/test/testfiles/tdset2.h5",
+  "hdf5/tools/test/testfiles/tdset_idx.h5",
+  "hdf5/tools/test/testfiles/tempty.h5",
+  "hdf5/tools/test/testfiles/textlink.h5",
+  "hdf5/tools/test/testfiles/textlinkfar.h5",
+  "hdf5/tools/test/testfiles/textlinksrc.h5",
+  "hdf5/tools/test/testfiles/textlinktar.h5",
+  "hdf5/tools/test/testfiles/textpfe.h5",
+  "hdf5/tools/test/testfiles/tfcontents1.h5",
+  "hdf5/tools/test/testfiles/tfcontents2.h5",
+  "hdf5/tools/test/testfiles/tfilters.h5",
+  "hdf5/tools/test/testfiles/tfloat16.h5",
+  "hdf5/tools/test/testfiles/tfloat16_be.h5",
+  "hdf5/tools/test/testfiles/tfloat4.h5",
+  "hdf5/tools/test/testfiles/tfloat6.h5",
+  "hdf5/tools/test/testfiles/tfloat8.h5",
+  "hdf5/tools/test/testfiles/tfloatsattrs.h5",
+  "hdf5/tools/test/testfiles/tfpformat.h5",
+  "hdf5/tools/test/testfiles/tfvalues.h5",
+  "hdf5/tools/test/testfiles/tgroup.h5",
+  "hdf5/tools/test/testfiles/tgrp_comments.h5",
+  "hdf5/tools/test/testfiles/tgrpnullspace.h5",
+  "hdf5/tools/test/testfiles/thlink.h5",
+  "hdf5/tools/test/testfiles/thyperslab.h5",
+  "hdf5/tools/test/testfiles/tintascii.h5",
+  "hdf5/tools/test/testfiles/tints4dims.h5",
+  "hdf5/tools/test/testfiles/tintsattrs.h5",
+  "hdf5/tools/test/testfiles/tintsnodata.h5",
+  "hdf5/tools/test/testfiles/tlarge_objname.h5",
+  "hdf5/tools/test/testfiles/tldouble.h5",
+  "hdf5/tools/test/testfiles/tldouble_scalar.h5",
+  "hdf5/tools/test/testfiles/tlonglinks.h5",
+  "hdf5/tools/test/testfiles/tloop.h5",
+  "hdf5/tools/test/testfiles/tnamed_dtype_attr.h5",
+  "hdf5/tools/test/testfiles/tnestedcmpddt.h5",
+  "hdf5/tools/test/testfiles/tnestedcomp.h5",
+  "hdf5/tools/test/testfiles/tno-subset.h5",
+  "hdf5/tools/test/testfiles/tnullspace.h5",
+  "hdf5/tools/test/testfiles/torderattr.h5",
+  "hdf5/tools/test/testfiles/tordergr.h5",
+  "hdf5/tools/test/testfiles/trefer_attr.h5",
+  "hdf5/tools/test/testfiles/trefer_compat.h5",
+  "hdf5/tools/test/testfiles/trefer_ext1.h5",
+  "hdf5/tools/test/testfiles/trefer_ext2.h5",
+  "hdf5/tools/test/testfiles/trefer_grp.h5",
+  "hdf5/tools/test/testfiles/trefer_obj.h5",
+  "hdf5/tools/test/testfiles/trefer_obj_del.h5",
+  "hdf5/tools/test/testfiles/trefer_param.h5",
+  "hdf5/tools/test/testfiles/trefer_reg.h5",
+  "hdf5/tools/test/testfiles/trefer_reg_1d.h5",
+  "hdf5/tools/test/testfiles/tsaf.h5",
+  "hdf5/tools/test/testfiles/tscalarattrintsize.h5",
+  "hdf5/tools/test/testfiles/tscalarintattrsize.h5",
+  "hdf5/tools/test/testfiles/tscalarintsize.h5",
+  "hdf5/tools/test/testfiles/tscalarstring.h5",
+  "hdf5/tools/test/testfiles/tslink.h5",
+  "hdf5/tools/test/testfiles/tsoftlinks.h5",
+  "hdf5/tools/test/testfiles/tst_onion_dset_1d.h5",
+  "hdf5/tools/test/testfiles/tst_onion_dset_ext.h5",
+  "hdf5/tools/test/testfiles/tst_onion_objs.h5",
+  "hdf5/tools/test/testfiles/tstr.h5",
+  "hdf5/tools/test/testfiles/tstr2.h5",
+  "hdf5/tools/test/testfiles/tstr3.h5",
+  "hdf5/tools/test/testfiles/tudfilter.h5",
+  "hdf5/tools/test/testfiles/tudfilter2.h5",
+  "hdf5/tools/test/testfiles/tudlink.h5",
+  "hdf5/tools/test/testfiles/tvldtypes1.h5",
+  "hdf5/tools/test/testfiles/tvldtypes2.h5",
+  "hdf5/tools/test/testfiles/tvldtypes3.h5",
+  "hdf5/tools/test/testfiles/tvldtypes4.h5",
+  "hdf5/tools/test/testfiles/tvldtypes5.h5",
+  "hdf5/tools/test/testfiles/tvlenstr_array.h5",
+  "hdf5/tools/test/testfiles/tvlstr.h5",
+  "hdf5/tools/test/testfiles/tvms.h5",
+  "hdf5/tools/test/testfiles/twithub.h5",
+  "hdf5/tools/test/testfiles/twithub513.h5",
+  "hdf5/tools/test/testfiles/txtfp32.h5",
+  "hdf5/tools/test/testfiles/txtfp64.h5",
+  "hdf5/tools/test/testfiles/txtin16.h5",
+  "hdf5/tools/test/testfiles/txtin32.h5",
+  "hdf5/tools/test/testfiles/txtin8.h5",
+  "hdf5/tools/test/testfiles/txtstr.h5",
+  "hdf5/tools/test/testfiles/txtuin16.h5",
+  "hdf5/tools/test/testfiles/txtuin32.h5",
+  "hdf5/tools/test/testfiles/vds/1_a.h5",
+  "hdf5/tools/test/testfiles/vds/1_b.h5",
+  "hdf5/tools/test/testfiles/vds/1_c.h5",
+  "hdf5/tools/test/testfiles/vds/1_d.h5",
+  "hdf5/tools/test/testfiles/vds/1_e.h5",
+  "hdf5/tools/test/testfiles/vds/1_f.h5",
+  "hdf5/tools/test/testfiles/vds/1_vds.h5",
+  "hdf5/tools/test/testfiles/vds/2_a.h5",
+  "hdf5/tools/test/testfiles/vds/2_b.h5",
+  "hdf5/tools/test/testfiles/vds/2_c.h5",
+  "hdf5/tools/test/testfiles/vds/2_d.h5",
+  "hdf5/tools/test/testfiles/vds/2_e.h5",
+  "hdf5/tools/test/testfiles/vds/2_vds.h5",
+  "hdf5/tools/test/testfiles/vds/3_1_vds.h5",
+  "hdf5/tools/test/testfiles/vds/3_2_vds.h5",
+  "hdf5/tools/test/testfiles/vds/4_0.h5",
+  "hdf5/tools/test/testfiles/vds/4_1.h5",
+  "hdf5/tools/test/testfiles/vds/4_2.h5",
+  "hdf5/tools/test/testfiles/vds/4_vds.h5",
+  "hdf5/tools/test/testfiles/vds/5_a.h5",
+  "hdf5/tools/test/testfiles/vds/5_b.h5",
+  "hdf5/tools/test/testfiles/vds/5_c.h5",
+  "hdf5/tools/test/testfiles/vds/5_vds.h5",
+  "hdf5/tools/test/testfiles/vds/a.h5",
+  "hdf5/tools/test/testfiles/vds/b.h5",
+  "hdf5/tools/test/testfiles/vds/c.h5",
+  "hdf5/tools/test/testfiles/vds/d.h5",
+  "hdf5/tools/test/testfiles/vds/f-0.h5",
+  "hdf5/tools/test/testfiles/vds/f-3.h5",
+  "hdf5/tools/test/testfiles/vds/vds-eiger.h5",
+  "hdf5/tools/test/testfiles/vds/vds-percival-unlim-maxmin.h5",
+  "hdf5/tools/test/testfiles/xml/tbitfields.h5",
+  "hdf5/tools/test/testfiles/xml/tcompound2.h5",
+  "hdf5/tools/test/testfiles/xml/tdset2.h5",
+  "hdf5/tools/test/testfiles/xml/tenum.h5",
+  "hdf5/tools/test/testfiles/xml/test35.nc",
+  "hdf5/tools/test/testfiles/xml/tloop2.h5",
+  "hdf5/tools/test/testfiles/xml/tmany.h5",
+  "hdf5/tools/test/testfiles/xml/tname-amp.h5",
+  "hdf5/tools/test/testfiles/xml/tname-apos.h5",
+  "hdf5/tools/test/testfiles/xml/tname-gt.h5",
+  "hdf5/tools/test/testfiles/xml/tname-lt.h5",
+  "hdf5/tools/test/testfiles/xml/tname-quot.h5",
+  "hdf5/tools/test/testfiles/xml/tname-sp.h5",
+  "hdf5/tools/test/testfiles/xml/tnodata.h5",
+  "hdf5/tools/test/testfiles/xml/tobjref.h5",
+  "hdf5/tools/test/testfiles/xml/topaque.h5",
+  "hdf5/tools/test/testfiles/xml/tref-escapes-at.h5",
+  "hdf5/tools/test/testfiles/xml/tref-escapes.h5",
+  "hdf5/tools/test/testfiles/xml/tref.h5",
+  "hdf5/tools/test/testfiles/xml/tstring-at.h5",
+  "hdf5/tools/test/testfiles/xml/tstring.h5",
+  "hdf5/tools/test/testfiles/zerodim.h5",
+  "netcdf-c/h5_test/ref_tst_h_compounds.h5",
+  "netcdf-c/h5_test/ref_tst_h_compounds2.h5",
+  "netcdf-c/nc_test4/ref_hdf5_compat1.nc",
+  "netcdf-c/nc_test4/ref_hdf5_compat2.nc",
+  "netcdf-c/nc_test4/ref_hdf5_compat3.nc",
+  "netcdf-c/nc_test4/ref_szip.h5",
+  "netcdf-c/nc_test4/ref_tst_compounds.nc",
+  "netcdf-c/nc_test4/ref_tst_dims.nc",
+  "netcdf-c/nc_test4/ref_tst_interops4.nc",
+  "netcdf-c/nc_test4/ref_tst_xplatform2_1.nc",
+  "netcdf-c/nc_test4/ref_tst_xplatform2_2.nc",
+  "netcdf-c/nc_test4/tdset.h5",
+  "netcdf-c/ncdump/ref_nc_test_netcdf4_4_0.nc",
+  "netcdf-c/ncdump/ref_no_ncproperty.nc",
+  "netcdf-c/ncdump/ref_provenance_v1.nc",
+  "netcdf-c/ncdump/ref_test_corrupt_magic.nc",
+  "netcdf-c/ncdump/ref_tst_compounds2.nc",
+  "netcdf-c/ncdump/ref_tst_compounds3.nc",
+  "netcdf-c/ncdump/ref_tst_compounds4.nc",
+  "netcdf-c/ncdump/ref_tst_irish_rover.nc",
+  "netcdf4-python/examples/data/prmsl.2000.nc",
+  "netcdf4-python/examples/data/prmsl.2001.nc",
+  "netcdf4-python/examples/data/prmsl.2002.nc",
+  "netcdf4-python/examples/data/prmsl.2003.nc",
+  "netcdf4-python/examples/data/prmsl.2004.nc",
+  "netcdf4-python/examples/data/prmsl.2005.nc",
+  "netcdf4-python/examples/data/prmsl.2006.nc",
+  "netcdf4-python/examples/data/prmsl.2007.nc",
+  "netcdf4-python/examples/data/prmsl.2008.nc",
+  "netcdf4-python/examples/data/prmsl.2009.nc",
+  "netcdf4-python/examples/data/prmsl.2010.nc",
+  "netcdf4-python/examples/data/prmsl.2011.nc",
+  "netcdf4-python/examples/data/rtofs_glo_3dz_f006_6hrly_reg3.nc",
+  "netcdf4-python/test/20171025_2056.Cloud_Top_Height.nc",
+  "netcdf4-python/test/issue1152.nc",
+  "netcdf4-python/test/issue671.nc",
+  "netcdf4-python/test/issue672.nc",
+  "netcdf4-python/test/test_gold.nc",
+  "usnistgov_h5wasm/test/array.h5",
+  "usnistgov_h5wasm/test/compressed.h5",
+  "usnistgov_h5wasm/test/empty.h5",
+  "usnistgov_h5wasm/test/float16.h5",
+  "usnistgov_h5wasm/test/vlen.h5",
+  "xarray-data/ROMS_example.nc",
+  "xarray-data/basin_mask.nc",
+  "xarray-data/imerghh_730.hdf5",
+  "xarray-data/precipitation.nc4"
+ ]
+}

conformance/check.py

@@ -0,0 +1,88 @@
+#!/usr/bin/env python3
+"""check.py <results_dir> <baseline.json> [--update]
+
+The conformance gate. Fails (exit 1) when
+  * clawhdf5 panicked, hung, crashed or ran out of memory on any file, or
+  * the ok count fell below the baseline's, or
+  * a file the baseline lists as ok is no longer ok (even if another file
+    became ok and the total held).
+New ok files are reported so the baseline can be raised (--update rewrites it
+from the results).
+"""
+import json
+import os
+import sys
+
+FATAL = ("panic", "hang", "crash", "oom")
+
+
+def main():
+    args = [a for a in sys.argv[1:] if not a.startswith("--")]
+    update = "--update" in sys.argv
+    res_dir, base_path = args
+    res = json.load(open(os.path.join(res_dir, "results.json")))
+    rows = res["rows"]
+    counts = {}
+    per_corpus = {}
+    for r in rows:
+        counts[r["class"]] = counts.get(r["class"], 0) + 1
+        pc = per_corpus.setdefault(r["corpus"], {})
+        pc[r["class"]] = pc.get(r["class"], 0) + 1
+    ok_files = sorted(r["file"] for r in rows if r["class"] == "ok")
+
+    if update:
+        meta = {}
+        mp = os.path.join(res_dir, "report-meta.json")
+        if os.path.exists(mp):
+            meta = json.load(open(mp))
+        base = {
+            "comment": "conformance/run.sh fails if the ok count drops below `ok` or a file in `ok_files` stops being ok. "
+                       "Regenerate with `conformance/run.sh --update-baseline` after an intended change.",
+            "commit": meta.get("commit", ""),
+            "date": meta.get("date", ""),
+            "reference": meta.get("reference", ""),
+            "files": len(rows),
+            "ok": len(ok_files),
+            "counts": dict(sorted(counts.items())),
+            "per_corpus": {k: dict(sorted(v.items())) for k, v in sorted(per_corpus.items())},
+            "ok_files": ok_files,
+        }
+        with open(base_path, "w") as fh:
+            json.dump(base, fh, indent=1)
+            fh.write("\n")
+        print(f"baseline updated: {len(ok_files)} ok of {len(rows)} files -> {base_path}")
+        return 0
+
+    base = json.load(open(base_path))
+    failures = []
+    fatal = [r for r in rows if r["class"] in FATAL]
+    for r in fatal:
+        failures.append(f"{r['class']}: {r['file']}: {r['ours_detail'][:200]}")
+    if len(ok_files) < base["ok"]:
+        failures.append(f"ok count dropped: {len(ok_files)} < baseline {base['ok']}")
+    now_ok = set(ok_files)
+    by_file = {r["file"]: r for r in rows}
+    for f in base["ok_files"]:
+        if f not in now_ok:
+            r = by_file.get(f)
+            why = f"now {r['class']}: {(r['ours_detail'] or r['first_issue'])[:200]}" if r else "no longer in the corpus"
+            failures.append(f"regressed: {f}: {why}")
+    gained = sorted(now_ok - set(base["ok_files"]))
+
+    print(f"conformance: {len(ok_files)} ok of {len(rows)} files (baseline {base['ok']} of {base['files']}); "
+          + ", ".join(f"{k} {v}" for k, v in sorted(counts.items())))
+    if gained:
+        print(f"{len(gained)} file(s) newly ok — raise the baseline with `conformance/run.sh --update-baseline`:")
+        for f in gained:
+            print(f"  + {f}")
+    if failures:
+        print(f"CONFORMANCE GATE FAILED ({len(failures)}):")
+        for f in failures:
+            print(f"  - {f}")
+        return 1
+    print("conformance gate passed")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

conformance/compare.py

@@ -0,0 +1,322 @@
+#!/usr/bin/env python3
+"""compare.py <results_dir>: classify each file and group failures by root cause.
+
+Writes <results_dir>/results.csv, results.json and summary.md.
+File classes (first match wins):
+  hang, oom, crash, panic       ours: timeout / allocation failure / signal / any panic (caught or not)
+  h5py-cannot-read              libhdf5/h5py failed to open the file (or crashed/hung)
+  ref-bug                       every issue is an object we refuse that h5py reads only through a
+                                libhdf5 bug, confirmed in this run by ref_bugs.py (its values
+                                change with the reading process's heap)
+  our-error                     we fail to open, list, or read something h5py reads
+  mismatch                      we read something with different shape/values, or a different object set
+  ok
+"""
+import collections
+import csv
+import json
+import os
+import re
+import sys
+
+R = sys.argv[1]
+RUNS = os.path.join(R, "runs")
+
+# Objects ref_bugs.py confirmed in this run: h5py's values for them come from
+# libhdf5 reading memory the file does not determine.
+try:
+    REF_BUGS = {(b["file"], b["object"])
+                for b in json.load(open(os.path.join(R, "ref_bugs.json")))["read_bugs"] if b.get("confirmed")}
+except (OSError, ValueError, KeyError):
+    REF_BUGS = set()
+
+
+def is_ref_bug(rel, issue):
+    """An our-error on reading an object that ref_bugs.py confirmed."""
+    kind, detail = issue[0], issue[1]
+    return kind == "our-error" and any(f == rel and detail.startswith(obj + ": error: ") for f, obj in REF_BUGS)
+
+
+def load(d, name):
+    rc_p = os.path.join(d, name + ".rc")
+    if not os.path.exists(rc_p):
+        return None
+    rc = int(open(rc_p).read().strip() or -1)
+    err = open(os.path.join(d, name + ".err"), errors="replace").read()
+    js = None
+    try:
+        js = json.load(open(os.path.join(d, name + ".json")))
+    except Exception:  # noqa: BLE001
+        pass
+    return {"rc": rc, "err": err, "json": js}
+
+
+def proc_status(p):
+    """-> (status, detail)"""
+    if p is None:
+        return "missing", ""
+    rc, err = p["rc"], p["err"]
+    first_panic = next((ln for ln in err.splitlines() if ln.startswith("PANIC:") or "panicked at" in ln), "")
+    if rc == 0 and p["json"] is not None:
+        return "ok", ""
+    if rc == 137 or rc == 124:
+        return "hang", f"timeout ({os.environ.get('TMO', '20')} s)"
+    if "memory allocation of" in err or "MemoryError" in err or "std::bad_alloc" in err:
+        m = re.search(r"memory allocation of \d+ bytes failed", err)
+        return "oom", m.group(0) if m else "allocation failure"
+    if "overflowed its stack" in err:
+        return "crash", "stack overflow"
+    if rc == 101:
+        return "panic", first_panic or (err.strip().splitlines() or [""])[-1]
+    if rc in (134, 139, 136, 135, 132) or rc > 128:
+        sig = {134: "SIGABRT", 139: "SIGSEGV", 136: "SIGFPE", 135: "SIGBUS", 132: "SIGILL"}.get(rc, f"signal {rc - 128}")
+        tail = [ln for ln in err.strip().splitlines() if ln.strip()][-1:]
+        return "crash", f"{sig}: {tail[0][:200] if tail else ''}"
+    tail = [ln for ln in err.strip().splitlines() if ln.strip()][-1:]
+    return "crash", f"rc={rc}: {tail[0][:200] if tail else ''}"
+
+
+def norm(msg):
+    m = msg.split("\n")[0]
+    m = re.sub(r"0x[0-9a-fA-F]+", "X", m)
+    m = re.sub(r'"[^"]*"', '"…"', m)
+    m = re.sub(r"'[^']*'", "'…'", m)
+    m = re.sub(r"\d+", "N", m)
+    return m[:160]
+
+
+def panic_head(msg):
+    """First line + first clawhdf5 frame of a PANIC record."""
+    lines = msg.split("\n")
+    frame = next((ln.strip() for ln in lines[1:] if "clawhdf5_format" in ln), "")
+    return lines[0][:300], frame[:300]
+
+
+def eq_shape(a, b):
+    return a == b
+
+
+rows = []
+issues_by_file = {}
+root_causes = collections.defaultdict(lambda: {"files": set(), "count": 0, "examples": []})
+mismatch_causes = collections.defaultdict(lambda: {"files": set(), "count": 0, "examples": []})
+panics = []
+ref_only_errors = collections.Counter()
+incomparable = collections.Counter()
+# Values ref.py corrected for a known h5py bug: (file, object, fixes, same as ours)
+ref_fixes = []
+
+
+def add(bucket, key, file, example):
+    b = bucket[key]
+    b["count"] += 1
+    if file not in b["files"] and len(b["examples"]) < 6:
+        b["examples"].append(example)
+    b["files"].add(file)
+
+
+files = [ln.strip() for ln in open(os.path.join(R, "files.txt")) if ln.strip()]
+for rel in files:
+    d = os.path.join(RUNS, rel.replace("/", "__"))
+    corpus = rel.split("/")[0]
+    ours, ref = load(d, "ours"), load(d, "ref")
+    h5dump = load(d, "h5dump")
+    os_, od = proc_status(ours)
+    rs, rd = proc_status(ref)
+    oj = ours["json"] if ours else None
+    rj = ref["json"] if ref else None
+    issues = []  # (kind, detail)
+    caught_panics = []
+
+    def scan_err(path, what, msg):
+        if msg.startswith("PANIC:"):
+            caught_panics.append((path, what, msg))
+
+    if oj:
+        for o in oj.get("objects", []):
+            for k in ("error", "attrs_error", "list_error"):
+                if k in o:
+                    scan_err(o["path"], k, o[k])
+            for an, av in (o.get("attrs") or {}).items():
+                if "error" in av:
+                    scan_err(o["path"], f"attr {an}", av["error"])
+        if oj.get("open_error", "").startswith("PANIC:"):
+            caught_panics.append(("<open>", "open", oj["open_error"]))
+
+    ref_open_fail = rs != "ok" or (rj is not None and "open_error" in rj)
+    ours_open_err = oj.get("open_error") if oj else None
+    n_obj = n_ok = 0
+    if os_ == "ok" and rj and not ref_open_fail and not ours_open_err:
+        ro = {x["path"]: x for x in rj.get("objects", [])}
+        oo = {x["path"]: x for x in oj.get("objects", [])}
+        our_list_errors = [x for x in oo.values() if "list_error" in x]
+        for p in sorted(set(ro) | set(oo)):
+            a, b = ro.get(p), oo.get(p)
+            n_obj += 1
+            if a is None:
+                issues.append(("mismatch", f"extra object {p} (kind={b.get('kind')})", "extra-object", b))
+                continue
+            if b is None:
+                if our_list_errors:
+                    continue  # accounted for by the list_error
+                issues.append(("mismatch", f"missing object {p} (kind={a.get('kind')})", "missing-object", a))
+                continue
+            ok = True
+            if a.get("kind") != b.get("kind") and "error" not in b and "error" not in a:
+                issues.append(("mismatch", f"{p}: kind {a.get('kind')} vs ours {b.get('kind')}", "kind", b))
+                ok = False
+            # h5py could not open the object at all: it read none of its
+            # attributes or links, so there is nothing to compare ours with
+            # (the object's own error is compared above and below).
+            ref_unopened = a.get("kind") == "unknown" and "error" in a
+            for k in ("error", "list_error", "attrs_error"):
+                if ref_unopened and k != "error":
+                    continue
+                if k in b and k not in a:
+                    issues.append(("our-error", f"{p}: {k}: {b[k]}", b[k], b))
+                    ok = False
+                elif k in a and k not in b and k == "error":
+                    ref_only_errors[norm(a[k])] += 1
+            if a.get("kind") == "dataset" and "error" not in a and "error" not in b:
+                if "skipped" in a or "skipped" in b:
+                    pass
+                elif a.get("converted"):
+                    incomparable[f"dataset {a['converted']}"] += 1
+                elif a.get("shape") != b.get("shape"):
+                    issues.append(("mismatch", f"{p}: shape {a.get('shape')} vs ours {b.get('shape')}", "shape", b))
+                    ok = False
+                elif a.get("hash") != b.get("hash"):
+                    issues.append(("mismatch", f"{p}: values differ (h5py {a.get('dtype')} vs ours {b.get('dtype')})", "values", b | {"ref_head": a.get("head"), "ref_dtype": a.get("dtype")}))
+                    ok = False
+                if a.get("ref_fix") and "hash" in b:
+                    ref_fixes.append((rel, p, a["ref_fix"], a.get("hash") == b.get("hash")))
+            ra, oa = a.get("attrs") or {}, b.get("attrs") or {}
+            if "attrs_error" not in b and "attrs_error" not in a and not ref_unopened:
+                for an in sorted(set(ra) | set(oa)):
+                    x, y = ra.get(an), oa.get(an)
+                    if x is None:
+                        issues.append(("mismatch", f"{p}@{an}: extra attribute", "extra-attr", y or {}))
+                    elif y is None:
+                        issues.append(("mismatch", f"{p}@{an}: missing attribute", "missing-attr", x))
+                    elif "error" in y and "error" not in x:
+                        issues.append(("our-error", f"{p}@{an}: {y['error']}", y["error"], y))
+                    elif "error" in x:
+                        continue
+                    elif x.get("converted"):
+                        incomparable[f"attr {x['converted']}"] += 1
+                    elif x.get("shape") != y.get("shape"):
+                        issues.append(("mismatch", f"{p}@{an}: attr shape {x.get('shape')} vs ours {y.get('shape')}", "attr-shape", y | {"ref_dtype": x.get("dtype")}))
+                    elif x.get("hash") != y.get("hash"):
+                        issues.append(("mismatch", f"{p}@{an}: attr values differ (h5py {x.get('dtype')} vs ours {y.get('dtype')})", "attr-values", y | {"ref_head": x.get("head"), "ref_dtype": x.get("dtype")}))
+                    if x.get("ref_fix") and "hash" in y:
+                        ref_fixes.append((rel, f"{p}@{an}", x["ref_fix"], x.get("hash") == y.get("hash")))
+            if ok:
+                n_ok += 1
+
+    # classify
+    if os_ in ("hang", "oom", "crash", "panic"):
+        cls = os_
+    elif caught_panics:
+        cls = "panic"
+    elif ref_open_fail:
+        cls = "h5py-cannot-read"
+    elif issues and all(is_ref_bug(rel, i) for i in issues):
+        cls = "ref-bug"
+    elif ours_open_err:
+        cls = "our-error"
+        issues.append(("our-error", f"open: {ours_open_err}", ours_open_err, {}))
+    elif any(i[0] == "our-error" for i in issues):
+        cls = "our-error"
+    elif issues:
+        cls = "mismatch"
+    else:
+        cls = "ok"
+
+    if os_ in ("hang", "oom", "crash", "panic") or caught_panics:
+        panics.append({
+            "file": rel, "class": cls, "detail": od,
+            "stderr": (ours["err"] if ours else "")[:3000],
+            "caught": [(p, w, m[:2500]) for p, w, m in caught_panics[:3]],
+            "n_caught": len(caught_panics),
+        })
+    # A ref-bug file's differences are listed with the evidence instead.
+    for kind, detail, key, rec in (issues if cls != "ref-bug" else []):
+        if kind == "our-error":
+            add(root_causes, norm(key), rel, detail[:300])
+        else:
+            if key in ("values", "attr-values", "shape", "attr-shape"):
+                mk = f"{key}: ours={rec.get('dtype')} h5py={rec.get('ref_dtype')} layout={rec.get('layout','-')} filters={rec.get('filters','-')}"
+            else:
+                mk = key
+            add(mismatch_causes, mk, rel, detail[:300] + (f" | ref_head={rec.get('ref_head')} our_head={rec.get('head')}" if rec.get("ref_head") else ""))
+    ref_detail = rd if rs != "ok" else ((rj or {}).get("open_error") or "")
+    h5d = ""
+    if h5dump:
+        rc = h5dump["rc"]
+        h5d = {0: "ok", 1: "error", 137: "hang", 124: "hang", 134: "SIGABRT", 139: "SIGSEGV", 136: "SIGFPE", 135: "SIGBUS"}.get(rc, f"rc={rc}")
+        if "memory allocation" in h5dump["err"] or "Cannot allocate" in h5dump["err"]:
+            h5d += "(oom)"
+    rows.append({
+        "file": rel, "corpus": corpus, "class": cls,
+        "ours": os_ if os_ != "ok" else ("open-error" if ours_open_err else ("panic" if caught_panics else "ok")),
+        "ours_detail": (od or ours_open_err or (caught_panics[0][2].split("\n")[0] if caught_panics else ""))[:300],
+        "ref": rs if rs != "ok" else ("open-error" if (rj or {}).get("open_error") else "ok"),
+        "ref_detail": ref_detail[:300],
+        "h5dump_1_14_6": h5d,
+        "h5dump_detail": ([ln for ln in h5dump["err"].splitlines() if ln.strip()][-1:] or [""])[0][:200] if h5dump else "",
+        "objects": n_obj, "objects_ok": n_ok,
+        "issues": len(issues), "first_issue": issues[0][1][:300] if issues else "",
+        "superblock": (oj or {}).get("superblock_version", ""),
+    })
+    # the first issues of each file, for report.py's known-cause matching
+    issues_by_file[rel] = [
+        {"kind": k, "key": key, "detail": det[:300], "ours_dtype": rec.get("dtype"), "ref_dtype": rec.get("ref_dtype")}
+        for k, det, key, rec in issues[:50]
+    ]
+
+with open(os.path.join(R, "results.csv"), "w", newline="") as fh:
+    w = csv.DictWriter(fh, fieldnames=list(rows[0].keys()))
+    w.writeheader()
+    w.writerows(rows)
+
+
+def ser(b):
+    return {k: {"files": len(v["files"]), "count": v["count"], "examples": v["examples"], "file_list": sorted(v["files"])} for k, v in sorted(b.items(), key=lambda kv: -len(kv[1]["files"]))}
+
+
+json.dump({"rows": rows, "issues": issues_by_file, "root_causes": ser(root_causes), "mismatch_causes": ser(mismatch_causes),
+           "panics": panics, "incomparable": incomparable.most_common(), "ref_only_errors": ref_only_errors.most_common(),
+           "ref_fixes": ref_fixes, "ref_bugs_confirmed": sorted(REF_BUGS)},
+          open(os.path.join(R, "results.json"), "w"), indent=1)
+
+classes = ["ok", "our-error", "mismatch", "h5py-cannot-read", "ref-bug", "hang", "panic", "crash", "oom"]
+by_corpus = collections.defaultdict(collections.Counter)
+for r in rows:
+    by_corpus[r["corpus"]][r["class"]] += 1
+    by_corpus["ALL"][r["class"]] += 1
+lines = ["# Conformance sweep summary", "", "| corpus | files | " + " | ".join(classes) + " |", "|---" * (len(classes) + 2) + "|"]
+for c in sorted(by_corpus, key=lambda k: (k == "ALL", k)):
+    cnt = by_corpus[c]
+    lines.append(f"| {c} | {sum(cnt.values())} | " + " | ".join(str(cnt.get(k, 0)) for k in classes) + " |")
+lines += ["", "## Panics / hangs / crashes / OOM", ""]
+for p in panics:
+    lines.append(f"- **{p['file']}** [{p['class']}] {p['detail']}")
+    for path, what, m in p["caught"][:1]:
+        lines.append("  ```\n  " + f"{path} ({what}): " + m.replace("\n", "\n  ")[:1500] + "\n  ```")
+    if not p["caught"] and p["stderr"]:
+        lines.append("  ```\n  " + p["stderr"].strip()[:1500].replace("\n", "\n  ") + "\n  ```")
+lines += ["", "## Our-error root causes (files affected)", ""]
+for k, v in ser(root_causes).items():
+    lines.append(f"- [{v['files']} files, {v['count']} objs] `{k}`")
+    for ex in v["examples"][:3]:
+        lines.append(f"    - {ex}")
+lines += ["", "## Mismatch root causes", ""]
+for k, v in ser(mismatch_causes).items():
+    lines.append(f"- [{v['files']} files, {v['count']} objs] `{k}`")
+    for ex in v["examples"][:3]:
+        lines.append(f"    - {ex}")
+lines += ["", "## Objects h5py fails on but we read (top)", ""]
+for k, n in ref_only_errors.most_common(15):
+    lines.append(f"- {n} x `{k}`")
+open(os.path.join(R, "summary.md"), "w").write("\n".join(lines) + "\n")
+print("\n".join(lines[:4 + len(by_corpus)]))

conformance/corpus.txt

@@ -0,0 +1,19 @@
+# Conformance corpora, pinned by commit. fetch-corpus.sh reads this file.
+#
+# name  git-url  commit  root  [sparse-checkout patterns...]
+#
+# `root` is the directory inside the checkout that is swept ("." = all of it).
+# Patterns are git non-cone sparse-checkout patterns; none = whole repository.
+# Every file under <root> with an HDF5/netCDF-4 extension is probed; for
+# cve_hdf5 the extension-less files in cvefiles/ and fuzzerfiles/ are too.
+# Licences: each corpus keeps its upstream licence; nothing here is committed
+# to this repository — the files are downloaded into the gitignored cache.
+hdf5              https://github.com/HDFGroup/hdf5.git            a3cf1ea82cc7a66e50029a688121e1b105a7ce88  .  *.h5 *.he5 *.nc *.hdf5 *.h5f
+cve_hdf5          https://github.com/HDFGroup/cve_hdf5.git        3fd1f5ae3869e01b8ae02b41d7108de7ffb1a374  .
+netcdf-c          https://github.com/Unidata/netcdf-c.git         beb7b9585273c1548386231a59b809d906359033  .  /nc_test4/*.nc /ncdump/*.nc /nc_test4/*.h5 /ncdump/*.h5 /h5_test/*.h5 /hdf5_test/*.h5
+NCAS-CMS_pyfive   https://github.com/NCAS-CMS/pyfive.git          8cf07b8749133f41c5e30b8a4c604486f687fe74  .  *.h5 *.hdf5 *.hdf *.nc *.he5
+usnistgov_h5wasm  https://github.com/usnistgov/h5wasm.git         02f6336527d2812783fcedabfbf42127ec8d06d2  .  *.h5 *.hdf5 *.hdf *.nc *.he5
+netcdf4-python    https://github.com/Unidata/netcdf4-python.git   6e67576d39aef8091fb20bd767b4f1a52ddc1bec  .  *.nc *.h5
+xarray-data       https://github.com/pydata/xarray-data.git       a35297e9da2cc99c811014f0c8a4297345a5c28d  .  /basin_mask.nc /precipitation.nc4 /imerghh_730.hdf5 /eraint_uvz.nc /ROMS_example.nc /tiny.nc
+# h5py 3.16.0 (tag 3.16.0), its test data files.
+h5py_data         https://github.com/h5py/h5py.git                b2f0347c4200333acd89b43733f1caa0c115162f  h5py/tests/data_files  /h5py/tests/data_files/*

conformance/fetch-corpus.sh

@@ -0,0 +1,39 @@
+#!/usr/bin/env bash
+# fetch-corpus.sh [cache_dir]
+#
+# Download the corpora pinned in conformance/corpus.txt into the (gitignored)
+# cache: <cache>/src/<name> is a shallow, sparse, blob-filtered checkout of the
+# pinned commit and <cache>/corpus/<name> links to the swept root inside it.
+# A corpus already checked out at its pinned commit is left alone, so a second
+# run costs nothing and needs no network.
+set -euo pipefail
+HERE="$(cd "$(dirname "$0")" && pwd)"
+CACHE="${1:-${CONFORMANCE_CACHE:-$HERE/.cache}}"
+mkdir -p "$CACHE/src" "$CACHE/corpus"
+CACHE="$(cd "$CACHE" && pwd)"
+
+retry() { local i; for i in 1 2 3 4; do "$@" && return 0; sleep $((i * 5)); done; return 1; }
+
+grep -v '^[[:space:]]*\(#\|$\)' "$HERE/corpus.txt" | while read -r name url commit root patterns; do
+  src="$CACHE/src/$name"
+  if [ -d "$src/.git" ] && [ "$(git -C "$src" rev-parse HEAD 2>/dev/null)" = "$commit" ]; then
+    echo "cached   $name @ ${commit:0:12}"
+  else
+    echo "fetching $name @ ${commit:0:12} from $url"
+    rm -rf "$src"
+    git init -q "$src"
+    git -C "$src" remote add origin "$url"
+    git -C "$src" config advice.detachedHead false
+    if [ -n "$patterns" ]; then
+      git -C "$src" config core.sparseCheckout true
+      # no-cone patterns (globs); `set -f` keeps the shell from expanding them
+      (set -f; printf '%s\n' $patterns) > "$src/.git/info/sparse-checkout"
+    fi
+    retry git -C "$src" fetch -q --depth 1 --filter=blob:none origin "$commit"
+    retry git -C "$src" checkout -q FETCH_HEAD
+    got="$(git -C "$src" rev-parse HEAD)"
+    [ "$got" = "$commit" ] || { echo "error: $name checked out $got, expected $commit" >&2; exit 1; }
+  fi
+  ln -sfn "$src/$root" "$CACHE/corpus/$name"
+done
+echo "corpus ready in $CACHE/corpus"

conformance/list_files.py

@@ -0,0 +1,48 @@
+#!/usr/bin/env python3
+"""list_files.py <corpus_dir>: print the files the sweep probes, one per line,
+as <corpus>/<path> in byte order.
+
+  * every file named *.h5 *.hdf5 *.he5 *.nc *.nc4 *.hdf *.h5f in each corpus,
+    except netCDF classic / 64-bit-offset / CDF5 files (magic "CDF"): they are
+    not HDF5, so neither side can read them and they say nothing;
+  * plus, for cve_hdf5, every file in cvefiles/ and fuzzerfiles/ except
+    .md/.c sources — the reproducers are mostly extension-less, and they are
+    kept whatever their bytes look like (that is their point).
+"""
+import os
+import sys
+
+EXTS = (".h5", ".hdf5", ".he5", ".nc", ".nc4", ".hdf", ".h5f")
+
+
+def walk(top):
+    for dirpath, dirnames, filenames in os.walk(top):
+        dirnames[:] = [d for d in dirnames if d != ".git"]
+        for fn in filenames:
+            p = os.path.join(dirpath, fn)
+            if os.path.isfile(p) and not os.path.islink(p):
+                yield os.path.relpath(p, top)
+
+
+def main(root):
+    out = set()
+    for corpus in sorted(os.listdir(root)):
+        top = os.path.join(root, corpus)
+        if not os.path.isdir(top):
+            continue
+        for rel in walk(top):
+            path = os.path.join(top, rel)
+            if rel.lower().endswith(EXTS):
+                with open(path, "rb") as fh:
+                    if fh.read(3) == b"CDF":
+                        continue
+                out.add(f"{corpus}/{rel}")
+            elif corpus == "cve_hdf5" and rel.split(os.sep)[0] in ("cvefiles", "fuzzerfiles") \
+                    and not rel.endswith((".md", ".c")):
+                out.add(f"{corpus}/{rel}")
+    for f in sorted(out, key=lambda s: s.encode()):
+        print(f)
+
+
+if __name__ == "__main__":
+    main(sys.argv[1])

conformance/probe/Cargo.lock

@@ -0,0 +1,492 @@
+# This file is automatically @generated by Cargo.
+# It is not intended for manual editing.
+version = 4
+
+[[package]]
+name = "adler2"
+version = "2.0.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "320119579fcad9c21884f5c4861d16174d0e06250625266f50fe6898340abefa"
+
+[[package]]
+name = "better_io"
+version = "0.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ef0a3155e943e341e557863e69a708999c94ede624e37865c8e2a91b94efa78f"
+
+[[package]]
+name = "block-buffer"
+version = "0.10.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3078c7629b62d3f0439517fa394996acacc5cbc91c5a20d8c658e77abd503a71"
+dependencies = [
+ "generic-array",
+]
+
+[[package]]
+name = "byteorder"
+version = "1.5.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1fd0f2584146f6f2ef48085050886acf353beff7305ebd1ae69500e27c67f64b"
+
+[[package]]
+name = "bzip2"
+version = "0.6.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f3a53fac24f34a81bc9954b5d6cfce0c21e18ec6959f44f56e8e90e4bb7c346c"
+dependencies = [
+ "libbz2-rs-sys",
+]
+
+[[package]]
+name = "cc"
+version = "1.5.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f360145194ee8e21db5ee7f3fcd4fe52210864c75c985dae33218202c8bbe040"
+dependencies = [
+ "find-msvc-tools",
+ "jobserver",
+ "libc",
+ "shlex",
+]
+
+[[package]]
+name = "cfg-if"
+version = "1.0.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "4e7648175b45a9a48536d676f68d918270699102aa8dab5496df06904c914600"
+
+[[package]]
+name = "clawhdf5-format"
+version = "2.7.0"
+dependencies = [
+ "byteorder",
+ "bzip2",
+ "flate2",
+ "libaec-sys",
+ "libc",
+ "lz4_flex",
+ "pco",
+ "portable-atomic",
+ "ruzstd",
+ "sha2",
+ "snap",
+ "zstd",
+]
+
+[[package]]
+name = "conformance-probe"
+version = "0.1.0"
+dependencies = [
+ "clawhdf5-format",
+ "serde_json",
+ "sha2",
+]
+
+[[package]]
+name = "cpufeatures"
+version = "0.2.17"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "59ed5838eebb26a2bb2e58f6d5b5316989ae9d08bab10e0e6d103e656d1b0280"
+dependencies = [
+ "libc",
+]
+
+[[package]]
+name = "crc32fast"
+version = "1.5.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "01a7799fd6b852db0e61728dde9a204c423b44d689dbd432522543614b490e78"
+dependencies = [
+ "cfg-if",
+]
+
+[[package]]
+name = "crunchy"
+version = "0.2.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "460fbee9c2c2f33933d720630a6a0bac33ba7053db5344fac858d4b8952d77d5"
+
+[[package]]
+name = "crypto-common"
+version = "0.1.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "78c8292055d1c1df0cce5d180393dc8cce0abec0a7102adb6c7b1eef6016d60a"
+dependencies = [
+ "generic-array",
+ "typenum",
+]
+
+[[package]]
+name = "digest"
+version = "0.10.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9ed9a281f7bc9b7576e61468ba615a66a5c8cfdff42420a70aa82701a3b1e292"
+dependencies = [
+ "block-buffer",
+ "crypto-common",
+]
+
+[[package]]
+name = "dtype_dispatch"
+version = "0.2.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ab23e69df104e2fd85ee63a533a22d2132ef5975dc6b36f9f3e5a7305e4a8ed7"
+
+[[package]]
+name = "find-msvc-tools"
+version = "0.1.14"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "aedcfb3409746eddb02b9e19ebda1c3394f759a152e48ee875a0844d1b955484"
+
+[[package]]
+name = "flate2"
+version = "1.1.10"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6e634e2e0ebac1ee034020da1ca582e17ffe4e0f5e985823721e168928136dcb"
+dependencies = [
+ "crc32fast",
+ "miniz_oxide",
+ "zlib-rs",
+]
+
+[[package]]
+name = "generic-array"
+version = "0.14.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "85649ca51fd72272d7821adaf274ad91c288277713d9c18820d8499a7ff69e9a"
+dependencies = [
+ "typenum",
+ "version_check",
+]
+
+[[package]]
+name = "getrandom"
+version = "0.4.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "300e883d756b2e4ec94e02791f39b04b522276138852cfc41d9fb7e904106099"
+dependencies = [
+ "cfg-if",
+ "libc",
+ "r-efi",
+]
+
+[[package]]
+name = "half"
+version = "2.7.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6ea2d84b969582b4b1864a92dc5d27cd2b77b622a8d79306834f1be5ba20d84b"
+dependencies = [
+ "cfg-if",
+ "crunchy",
+ "zerocopy",
+]
+
+[[package]]
+name = "itoa"
+version = "1.0.18"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8f42a60cbdf9a97f5d2305f08a87dc4e09308d1276d28c869c684d7777685682"
+
+[[package]]
+name = "jobserver"
+version = "0.1.35"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1c00acbd29eabad4a2392fa0e921c874934dbbf4194312ad20f04a0ed67a3cb3"
+dependencies = [
+ "getrandom",
+ "libc",
+]
+
+[[package]]
+name = "libaec-sys"
+version = "0.1.0"
+dependencies = [
+ "pkg-config",
+]
+
+[[package]]
+name = "libbz2-rs-sys"
+version = "0.2.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "34b357333733e8260735ba5894eb928c02ecc69c78715f01a8019e7fa7f2db4c"
+
+[[package]]
+name = "libc"
+version = "0.2.189"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3eaf3ede3fee6db1a4c2ee091bf8a8b4dccdc6d17f656fb07896ee72867612f2"
+
+[[package]]
+name = "lz4_flex"
+version = "0.11.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "373f5eceeeab7925e0c1098212f2fbc4d416adec9d35051a6ab251e824c1854a"
+dependencies = [
+ "twox-hash",
+]
+
+[[package]]
+name = "memchr"
+version = "2.8.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "cf8baf1c55e62ffcace7a9f06f4bd9cd3f0c4beb022d3b367256b91b87513d98"
+
+[[package]]
+name = "miniz_oxide"
+version = "0.9.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b63fbc4a50860e98e7b2aa7804ded1db5cbc3aff9193adaff57a6931bf7c4b4c"
+dependencies = [
+ "adler2",
+ "simd-adler32",
+]
+
+[[package]]
+name = "pco"
+version = "1.0.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "386342cad4c6e97f081568e5d910ea7d871314c843aa8fc564f2a6b64cab9456"
+dependencies = [
+ "better_io",
+ "dtype_dispatch",
+ "half",
+ "rand_xoshiro",
+]
+
+[[package]]
+name = "pkg-config"
+version = "0.3.34"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f6b464fbc74e149a392436b17d523f769e057cb6877f6a5c4618bc6f11800548"
+
+[[package]]
+name = "portable-atomic"
+version = "1.15.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "05c8b63e8d9609db387f0324918f81d68fe27748f084ef092fb35954d0539a85"
+
+[[package]]
+name = "proc-macro2"
+version = "1.0.107"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "985e7ec9bb745e6ce6535b544d84d6cd6f7ad8bd711c398938ae983b91a766d9"
+dependencies = [
+ "unicode-ident",
+]
+
+[[package]]
+name = "quote"
+version = "1.0.47"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1fbf4db142a473a8d80c26bbf18454ed458bf8d26c8219c331daecfdbd079001"
+dependencies = [
+ "proc-macro2",
+]
+
+[[package]]
+name = "r-efi"
+version = "6.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f8dcc9c7d52a811697d2151c701e0d08956f92b0e24136cf4cf27b57a6a0d9bf"
+
+[[package]]
+name = "rand_core"
+version = "0.6.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ec0be4795e2f6a28069bec0b5ff3e2ac9bafc99e6a9a7dc3547996c5c816922c"
+
+[[package]]
+name = "rand_xoshiro"
+version = "0.6.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6f97cdb2a36ed4183de61b2f824cc45c9f1037f28afe0a322e9fff4c108b5aaa"
+dependencies = [
+ "rand_core",
+]
+
+[[package]]
+name = "ruzstd"
+version = "0.9.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a252f5e20f038fe7b4ea53e073e65398d652c864cc162fc77c56c2f13717b888"
+dependencies = [
+ "twox-hash",
+]
+
+[[package]]
+name = "serde"
+version = "1.0.229"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "4148590afebada386688f18773da617792bf2ef03ffc1e4cbd2b1d45b023e0ba"
+dependencies = [
+ "serde_core",
+]
+
+[[package]]
+name = "serde_core"
+version = "1.0.229"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "67dca2c9c51e58a4791a4b1ed58308b39c64224d349a935ab5039aa360942a48"
+dependencies = [
+ "serde_derive",
+]
+
+[[package]]
+name = "serde_derive"
+version = "1.0.229"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e7a5d71263a5a7d47b41f6b3f06ba276f10cc18b0931f1799f710578e2309348"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 3.0.6",
+]
+
+[[package]]
+name = "serde_json"
+version = "1.0.151"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c841b55ecdae098c80dcae9cf767f6f8a0c2cdb3416bbef72181df4d0fe73f14"
+dependencies = [
+ "itoa",
+ "memchr",
+ "serde",
+ "serde_core",
+ "zmij",
+]
+
+[[package]]
+name = "sha2"
+version = "0.10.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a7507d819769d01a365ab707794a4084392c824f54a7a6a7862f8c3d0892b283"
+dependencies = [
+ "cfg-if",
+ "cpufeatures",
+ "digest",
+]
+
+[[package]]
+name = "shlex"
+version = "2.0.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f8fadd59c855ef2080decdef8ff161eb6661b86933c9d82e5ba29dc602a55aba"
+
+[[package]]
+name = "simd-adler32"
+version = "0.3.10"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3a219298ac11a56ea9a6d2120044824d6f01aeb034955e7af7bc16858527deea"
+
+[[package]]
+name = "snap"
+version = "1.1.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "199905e6153d6405f9728fe44daace35f8f837bbf830bb6e85fbd5828709a886"
+
+[[package]]
+name = "syn"
+version = "2.0.119"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "872831b642d1a07999a962a351ed35b955ea2cfc8f3862091e2a240a84f17297"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "unicode-ident",
+]
+
+[[package]]
+name = "syn"
+version = "3.0.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8593e8e72159ed2257d083c7a454a85cbf854f37a0966d8d483aff8c8a3ebcee"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "unicode-ident",
+]
+
+[[package]]
+name = "twox-hash"
+version = "2.1.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5283634e518fe9e82c7b20520bb4bc209009fd16c82077c802f8111ecbb0117a"
+
+[[package]]
+name = "typenum"
+version = "1.20.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b6f5e870be6c3b371b77fe0ee0bafb859fa4964b4404c27de1d380043c4dda20"
+
+[[package]]
+name = "unicode-ident"
+version = "1.0.26"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d245f478577f809a851594d02313b640fb437e0bb33866753cff937863096954"
+
+[[package]]
+name = "version_check"
+version = "0.9.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0b928f33d975fc6ad9f86c8f283853ad26bdd5b10b7f1542aa2fa15e2289105a"
+
+[[package]]
+name = "zerocopy"
+version = "0.8.59"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6df92bf3d9227be3d53173901ddbffac2babc27ae50f397776ffd6dc33f800cb"
+dependencies = [
+ "zerocopy-derive",
+]
+
+[[package]]
+name = "zerocopy-derive"
+version = "0.8.59"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ac4f328cf2f05d084e496c3e9c3f33ed0a183656a16e1fcec4d464d8373aec82"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.119",
+]
+
+[[package]]
+name = "zlib-rs"
+version = "0.6.8"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b268e58e7c693d7c271f93ffc4ba3b380412554231c85bf61ca7af91042a4112"
+
+[[package]]
+name = "zmij"
+version = "1.0.23"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "29666d0abbfad1e3dc4dcf6144730dd3a3ab225bbbdac83319345b1b44ccfc1b"
+
+[[package]]
+name = "zstd"
+version = "0.13.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e91ee311a569c327171651566e07972200e76fcfe2242a4fa446149a3881c08a"
+dependencies = [
+ "zstd-safe",
+]
+
+[[package]]
+name = "zstd-safe"
+version = "7.3.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "64d80649ab6db9d9f6f9c80a40becd948eda4714a0a5ac8c4d157a32231c7882"
+dependencies = [
+ "zstd-sys",
+]
+
+[[package]]
+name = "zstd-sys"
+version = "2.1.0+zstd.1.5.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0ef0a8027ec3ee71300ab3bcbcd0393f434aa72b91ca6d635a39941deae8eea0"
+dependencies = [
+ "cc",
+ "pkg-config",
+]

conformance/probe/Cargo.toml

@@ -0,0 +1,25 @@
+[package]
+name = "conformance-probe"
+version = "0.1.0"
+edition = "2024"
+rust-version = "1.92"
+publish = false
+description = "Walks an HDF5 file with clawhdf5-format and prints a canonical JSON description (see conformance/README.md)"
+
+# Deliberately outside the main workspace: `cargo test --workspace` never
+# builds it, and it links the optional C codecs (zstd, libaec) that the core
+# crates' default build must not.
+[workspace]
+
+[dependencies]
+clawhdf5-format = { path = "../../crates/clawhdf5-format", features = ["lz4", "zstd", "szip", "pcodec", "plugin-filters"] }
+serde_json = "1"
+sha2 = "0.10"
+
+[profile.release]
+# Keep panics catchable (the probe records them per object) and turn integer
+# overflow into a reported panic instead of silent wraparound.
+debug = 1
+overflow-checks = true
+debug-assertions = true
+panic = "unwind"

conformance/probe/src/main.rs

@@ -0,0 +1,960 @@
+//! Conformance probe: walks an HDF5 file with clawhdf5-format (the same calls
+//! the `clawhdf5` facade makes) and prints a canonical JSON description:
+//! every hard-linked object (sorted-name DFS, deduplicated by header address),
+//! and for each dataset / attribute its shape plus the SHA-256 of its values
+//! in a canonical encoding shared with `ref.py`.
+//!
+//! Canonical value encoding (per element, concatenated, row-major):
+//!   int / float / bitfield / enum / time : element bytes, little-endian
+//!   non-IEEE-layout float (e.g. N-Bit)   : the IEEE float of the same size it converts to
+//!   int with bit offset / short precision: the full-width integer it converts to
+//!   opaque                               : raw bytes
+//!   compound                             : members in declaration order (padding dropped)
+//!   array                                : base elements row-major
+//!   string (fixed or VL)                 : b'S' + u32le len + bytes (cut at first NUL, trailing spaces stripped)
+//!   VL sequence                          : b'V' + u32le count + base elements
+//!   reference                            : b'R' (payload not compared)
+//!
+//! Every object is processed inside catch_unwind; a caught panic is recorded
+//! with its message, location and the clawhdf5 frames of its backtrace.
+
+use std::cell::RefCell;
+use std::collections::HashSet;
+use std::panic::{self, AssertUnwindSafe};
+
+use clawhdf5_format::attribute::extract_attributes_full;
+use clawhdf5_format::data_layout::DataLayout;
+use clawhdf5_format::data_read;
+use clawhdf5_format::dataspace::{Dataspace, DataspaceType};
+use clawhdf5_format::datatype::{Datatype, DatatypeByteOrder};
+use clawhdf5_format::filter_pipeline::FilterPipeline;
+use clawhdf5_format::group_v1::{self, GroupEntry};
+use clawhdf5_format::group_v2;
+use clawhdf5_format::message_type::MessageType;
+use clawhdf5_format::object_header::{ObjectClass, ObjectHeader};
+use clawhdf5_format::signature;
+use clawhdf5_format::superblock::Superblock;
+use clawhdf5_format::symbol_table::SymbolTableMessage;
+use clawhdf5_format::vl_data::{VlResolver, check_element_size};
+use serde_json::{Map, Value, json};
+use sha2::{Digest, Sha256};
+
+const MAX_BYTES: u64 = 200 * 1024 * 1024;
+const MAX_OBJECTS: usize = 200_000;
+
+thread_local! {
+    static LAST_PANIC: RefCell<Option<String>> = const { RefCell::new(None) };
+}
+
+fn install_hook() {
+    panic::set_hook(Box::new(|info| {
+        let msg = if let Some(s) = info.payload().downcast_ref::<&str>() {
+            s.to_string()
+        } else if let Some(s) = info.payload().downcast_ref::<String>() {
+            s.clone()
+        } else {
+            "<non-string panic>".into()
+        };
+        let loc = info
+            .location()
+            .map(|l| format!("{}:{}", l.file(), l.line()))
+            .unwrap_or_default();
+        let bt = std::backtrace::Backtrace::force_capture().to_string();
+        // keep only frames from clawhdf5 code
+        let mut frames = Vec::new();
+        let lines: Vec<&str> = bt.lines().collect();
+        for (i, l) in lines.iter().enumerate() {
+            let t = l.trim();
+            if t.contains("clawhdf5_format::") || t.contains("conformance_probe::") {
+                let at = lines
+                    .get(i + 1)
+                    .map(|n| n.trim())
+                    .filter(|n| n.starts_with("at "))
+                    .map(|n| {
+                        let n = n.trim_start_matches("at ");
+                        match n.find("/crates/") {
+                            Some(p) => n[p + 1..].to_string(),
+                            None => n.to_string(),
+                        }
+                    })
+                    .unwrap_or_default();
+                let name = t.split_once(": ").map(|x| x.1).unwrap_or(t);
+                frames.push(format!("{name} ({at})"));
+                if frames.len() >= 12 {
+                    break;
+                }
+            }
+        }
+        let full = format!("PANIC: {msg} @ {loc}\n  {}", frames.join("\n  "));
+        eprintln!("{full}");
+        LAST_PANIC.with(|p| *p.borrow_mut() = Some(full));
+    }));
+}
+
+/// Run `f`, turning a panic into Err("PANIC: ...").
+fn guarded<T>(f: impl FnOnce() -> Result<T, String>) -> Result<T, String> {
+    match panic::catch_unwind(AssertUnwindSafe(f)) {
+        Ok(r) => r,
+        Err(_) => Err(LAST_PANIC
+            .with(|p| p.borrow_mut().take())
+            .unwrap_or_else(|| "PANIC: <unknown>".into())),
+    }
+}
+
+fn e<E: std::fmt::Debug>(x: E) -> String {
+    format!("{x:?}")
+}
+
+struct Ctx<'a> {
+    data: &'a [u8],
+    os: u8,
+    ls: u8,
+    base_dir: std::path::PathBuf,
+    /// Resolves variable-length elements as the library does (null
+    /// elements, strings cut at a NUL, heap objects of the wrong size
+    /// refused), caching each heap collection.
+    vl: RefCell<VlResolver<'a>>,
+}
+
+impl<'a> Ctx<'a> {
+    fn header(&self, addr: u64) -> Result<ObjectHeader, String> {
+        ObjectHeader::parse(self.data, addr as usize, self.os, self.ls).map_err(e)
+    }
+
+    fn payload(&self, h: &ObjectHeader, t: MessageType) -> Result<Option<Vec<u8>>, String> {
+        match h.messages.iter().find(|m| m.msg_type == t) {
+            None => Ok(None),
+            Some(m) => {
+                clawhdf5_format::shared_message::message_data(self.data, m, self.os, self.ls)
+                    .map(|c| Some(c.into_owned()))
+                    .map_err(e)
+            }
+        }
+    }
+
+    fn canon(&self, dt: &Datatype, b: &[u8], out: &mut Vec<u8>) -> Result<(), String> {
+        let size = dt.type_size() as usize;
+        if b.len() < size {
+            return Err(format!(
+                "canon: element slice {} < type size {size}",
+                b.len()
+            ));
+        }
+        match dt {
+            Datatype::FloatingPoint { .. } if !ieee_layout(dt) => {
+                canon_custom_float(dt, &b[..size], out)?
+            }
+            Datatype::FixedPoint { .. } if partial_int(dt) => {
+                canon_partial_int(dt, &b[..size], out)?
+            }
+            Datatype::FixedPoint { byte_order, .. }
+            | Datatype::BitField { byte_order, .. }
+            | Datatype::FloatingPoint { byte_order, .. } => match byte_order {
+                DatatypeByteOrder::LittleEndian => out.extend_from_slice(&b[..size]),
+                DatatypeByteOrder::BigEndian => out.extend(b[..size].iter().rev()),
+                DatatypeByteOrder::Vax => return Err("canon: VAX byte order".into()),
+            },
+            Datatype::Time { .. } | Datatype::Opaque { .. } => out.extend_from_slice(&b[..size]),
+            Datatype::String { .. } => canon_str(&b[..size], out),
+            Datatype::Compound { members, .. } => {
+                for m in members {
+                    let off = m.byte_offset as usize;
+                    let ms = m.datatype.type_size() as usize;
+                    if off.checked_add(ms).is_none_or(|end| end > size) {
+                        return Err(format!("canon: member {} out of bounds", m.name));
+                    }
+                    self.canon(&m.datatype, &b[off..off + ms], out)?;
+                }
+            }
+            Datatype::Reference { .. } => out.push(b'R'),
+            Datatype::Enumeration { base_type, .. } => self.canon(base_type, b, out)?,
+            Datatype::Array {
+                base_type,
+                dimensions,
+            } => {
+                let n: usize = dimensions.iter().map(|d| *d as usize).product();
+                let bs = base_type.type_size() as usize;
+                for i in 0..n {
+                    self.canon(base_type, &b[i * bs..], out)?;
+                }
+            }
+            Datatype::VariableLength {
+                size: vl_size,
+                is_string,
+                base_type,
+                ..
+            } => {
+                check_element_size(*vl_size, self.os).map_err(e)?;
+                let el = &b[..size];
+                if *is_string {
+                    let s = self.vl.borrow_mut().string_bytes(el).map_err(e)?;
+                    canon_str(&s[0], out);
+                } else {
+                    let bs = base_type.type_size() as usize;
+                    // The borrow ends here: the base type may itself be
+                    // variable-length.
+                    let seq = self.vl.borrow_mut().sequences(el, bs).map_err(e)?;
+                    let seq = &seq[0];
+                    let len = seq.len() / bs;
+                    out.push(b'V');
+                    out.extend_from_slice(&(len as u32).to_le_bytes());
+                    for i in 0..len {
+                        self.canon(base_type, &seq[i * bs..], out)?;
+                    }
+                }
+            }
+        }
+        Ok(())
+    }
+
+    /// Returns (shape json, n_elements)
+    fn shape(ds: &Dataspace) -> (Value, u64) {
+        match ds.space_type {
+            DataspaceType::Null => (Value::String("null".into()), 0),
+            DataspaceType::Scalar => (json!([]), 1),
+            DataspaceType::Simple => {
+                let n = ds.dimensions.iter().fold(1u64, |a, d| a.saturating_mul(*d));
+                (json!(ds.dimensions), n)
+            }
+        }
+    }
+
+    fn hash_values(
+        &self,
+        dt: &Datatype,
+        raw: &[u8],
+        n: u64,
+        rec: &mut Map<String, Value>,
+    ) -> Result<(), String> {
+        let size = dt.type_size() as usize;
+        let need = (n as usize).checked_mul(size).ok_or("n*size overflow")?;
+        if raw.len() != need {
+            return Err(format!(
+                "raw length {} != n_elements {n} * type_size {size}",
+                raw.len()
+            ));
+        }
+        let mut canon = Vec::with_capacity(need);
+        for i in 0..n as usize {
+            self.canon(dt, &raw[i * size..(i + 1) * size], &mut canon)?;
+        }
+        let h = Sha256::digest(&canon);
+        rec.insert("hash".into(), Value::String(hex(&h)));
+        rec.insert(
+            "head".into(),
+            Value::String(hex(&canon[..canon.len().min(48)])),
+        );
+        Ok(())
+    }
+
+    /// VDS source files resolve next to the virtual file; like the library,
+    /// refuse absolute paths and `..`.
+    fn vds_resolver(
+        &self,
+    ) -> impl Fn(&str) -> Result<Option<Vec<u8>>, clawhdf5_format::error::FormatError> + use<> {
+        let base = self.base_dir.clone();
+        move |name: &str| {
+            use clawhdf5_format::error::FormatError;
+            let p = std::path::Path::new(name);
+            if p.is_absolute()
+                || p.components()
+                    .any(|c| matches!(c, std::path::Component::ParentDir))
+            {
+                return Err(FormatError::ChunkedReadError(format!("refused {name}")));
+            }
+            match std::fs::read(base.join(p)) {
+                Ok(b) => Ok(Some(b)),
+                Err(err) if err.kind() == std::io::ErrorKind::NotFound => Ok(None),
+                Err(err) => Err(FormatError::ChunkedReadError(err.to_string())),
+            }
+        }
+    }
+
+    fn read_named_datatype(&self, h: &ObjectHeader) -> Result<(), String> {
+        let dtb = self
+            .payload(h, MessageType::Datatype)?
+            .ok_or("MissingMessage(Datatype)")?;
+        Datatype::parse_in_header(&dtb, h.version).map_err(e)?;
+        Ok(())
+    }
+
+    fn read_dataset(&self, h: &ObjectHeader, rec: &mut Map<String, Value>) -> Result<(), String> {
+        let dtb = self
+            .payload(h, MessageType::Datatype)?
+            .ok_or("MissingMessage(Datatype)")?;
+        let (dt, _) = Datatype::parse_in_header(&dtb, h.version).map_err(e)?;
+        rec.insert("dtype".into(), Value::String(dtype_str(&dt)));
+        let dsb = self
+            .payload(h, MessageType::Dataspace)?
+            .ok_or("MissingMessage(Dataspace)")?;
+        let mut ds = Dataspace::parse(&dsb, self.ls).map_err(e)?;
+        // A virtual dataset's extent can come from its sources (unlimited /
+        // printf mappings), as h5py reports it, rather than the stored one.
+        if let Some(lm) = h
+            .messages
+            .iter()
+            .find(|m| m.msg_type == MessageType::DataLayout)
+            && let Ok(dl @ DataLayout::Virtual { .. }) =
+                DataLayout::parse(&lm.data, self.os, self.ls)
+        {
+            let resolver = self.vds_resolver();
+            ds.dimensions = clawhdf5_format::vds::virtual_dataset_extent(
+                self.data,
+                &dl,
+                &ds,
+                self.os,
+                self.ls,
+                Some(&resolver),
+            )
+            .map_err(e)?;
+        }
+        let lm = h
+            .messages
+            .iter()
+            .find(|m| m.msg_type == MessageType::DataLayout)
+            .ok_or("MissingMessage(DataLayout)")?;
+        let dl = DataLayout::parse(&lm.data, self.os, self.ls).map_err(e)?;
+        // What libhdf5 checks when it opens the dataset (as File::dataset).
+        data_read::check_dataset_storage(&dl, &ds, &dt, self.data.len() as u64).map_err(e)?;
+        let (shape, n) = Self::shape(&ds);
+        rec.insert("shape".into(), shape);
+        if n.saturating_mul(dt.type_size() as u64) > MAX_BYTES {
+            rec.insert("skipped".into(), Value::String("too large".into()));
+            return Ok(());
+        }
+        rec.insert(
+            "layout".into(),
+            Value::String(
+                match &dl {
+                    DataLayout::Compact { .. } => "compact",
+                    DataLayout::Contiguous { .. } => "contiguous",
+                    DataLayout::Chunked { .. } => "chunked",
+                    DataLayout::Virtual { .. } => "virtual",
+                }
+                .into(),
+            ),
+        );
+        let pipeline = match self.payload(h, MessageType::FilterPipeline)? {
+            Some(p) => Some(FilterPipeline::parse(&p).map_err(e)?),
+            None => None,
+        };
+        if let Some(p) = &pipeline {
+            rec.insert(
+                "filters".into(),
+                json!(p.filters.iter().map(|f| f.filter_id).collect::<Vec<_>>()),
+            );
+        }
+        let raw = if matches!(dl, DataLayout::Virtual { .. }) {
+            let resolver = self.vds_resolver();
+            let fill = clawhdf5_format::fill_value::dataset_fill_value_in(
+                self.data,
+                &h.messages,
+                self.os,
+                self.ls,
+            )
+            .map_err(e)?;
+            clawhdf5_format::vds::read_virtual_dataset(
+                self.data,
+                &dl,
+                &ds,
+                &dt,
+                fill.as_deref(),
+                self.os,
+                self.ls,
+                Some(&resolver),
+            )
+            .map_err(e)?
+            .data
+        } else {
+            let cache = clawhdf5_format::chunk_cache::ChunkCache::new();
+            clawhdf5_format::fill_value::read_full_with_fill::<clawhdf5_format::error::FormatError>(
+                &h.messages,
+                self.data,
+                &dl,
+                &ds,
+                dt.type_size() as usize,
+                self.os,
+                self.ls,
+                || {
+                    data_read::read_raw_data_cached(
+                        self.data,
+                        &dl,
+                        &ds,
+                        &dt,
+                        pipeline.as_ref(),
+                        self.os,
+                        self.ls,
+                        &cache,
+                    )
+                },
+            )
+            .map_err(e)?
+        };
+        self.hash_values(&dt, &raw, n, rec)
+    }
+
+    fn attrs(&self, h: &ObjectHeader) -> Result<Map<String, Value>, String> {
+        let msgs = extract_attributes_full(self.data, h, self.os, self.ls).map_err(e)?;
+        let mut out = Map::new();
+        for a in &msgs {
+            let r = guarded(|| {
+                let mut rec = Map::new();
+                rec.insert("dtype".into(), Value::String(dtype_str(&a.datatype)));
+                let (shape, n) = Self::shape(&a.dataspace);
+                rec.insert("shape".into(), shape);
+                self.hash_values(&a.datatype, &a.raw_data, n, &mut rec)?;
+                Ok(rec)
+            });
+            let v = match r {
+                Ok(rec) => Value::Object(rec),
+                Err(msg) => json!({ "error": msg }),
+            };
+            out.insert(a.name.clone(), v);
+        }
+        Ok(out)
+    }
+
+    fn entries(&self, h: &ObjectHeader) -> Result<Vec<GroupEntry>, String> {
+        let v1 = h
+            .messages
+            .iter()
+            .find(|m| m.msg_type == MessageType::SymbolTable);
+        if let Some(m) = v1 {
+            let stm = SymbolTableMessage::parse(&m.data, self.os).map_err(e)?;
+            group_v1::resolve_v1_group_entries(self.data, &stm, self.os, self.ls).map_err(e)
+        } else if h
+            .messages
+            .iter()
+            .any(|m| m.msg_type == MessageType::LinkInfo || m.msg_type == MessageType::Link)
+        {
+            group_v2::resolve_v2_group_entries(self.data, h, self.os, self.ls).map_err(e)
+        } else {
+            Ok(Vec::new())
+        }
+    }
+}
+
+/// Element bytes as an unsigned integer (at most 16 bytes), honouring byte order.
+fn element_bits(b: &[u8], byte_order: &DatatypeByteOrder) -> Result<u128, String> {
+    if b.len() > 16 {
+        return Err(format!("canon: {}-byte numeric element", b.len()));
+    }
+    let mut v = 0u128;
+    match byte_order {
+        DatatypeByteOrder::LittleEndian => {
+            for (i, x) in b.iter().enumerate() {
+                v |= u128::from(*x) << (8 * i);
+            }
+        }
+        DatatypeByteOrder::BigEndian => {
+            for x in b {
+                v = (v << 8) | u128::from(*x);
+            }
+        }
+        DatatypeByteOrder::Vax => return Err("canon: VAX byte order".into()),
+    }
+    Ok(v)
+}
+
+fn field(v: u128, pos: u32, len: u32) -> u128 {
+    if len == 0 || pos >= 128 {
+        return 0;
+    }
+    let v = v >> pos;
+    if len >= 128 {
+        v
+    } else {
+        v & ((1u128 << len) - 1)
+    }
+}
+
+/// True when a float's bit fields are exactly IEEE 754 binary16/32/64 for its
+/// size. h5py hands back such a type's bytes untouched; any other layout (an
+/// N-Bit `H5Tset_precision` float, say) is *converted* by libhdf5 into the
+/// numpy float of the same size, so comparing raw bytes would be meaningless.
+fn ieee_layout(dt: &Datatype) -> bool {
+    let Datatype::FloatingPoint {
+        size,
+        bit_offset,
+        bit_precision,
+        exponent_location,
+        exponent_size,
+        mantissa_location,
+        mantissa_size,
+        exponent_bias,
+        ..
+    } = dt
+    else {
+        return true;
+    };
+    let std = match size {
+        2 => (16, 10, 5, 10, 15),
+        4 => (32, 23, 8, 23, 127),
+        8 => (64, 52, 11, 52, 1023),
+        _ => return true, // no same-size numpy float to convert to: compare raw
+    };
+    *bit_offset == 0
+        && (
+            *bit_precision,
+            *exponent_location,
+            *exponent_size,
+            *mantissa_size,
+            *exponent_bias,
+        ) == (std.0, std.1, std.2, std.3, std.4)
+        && *mantissa_location == 0
+}
+
+/// Canonicalise a non-IEEE-layout float the way libhdf5's float->float
+/// conversion presents it to h5py: as the IEEE float of the same size.
+/// Assumes the implied-leading-one normalisation and the sign bit at the top
+/// of the precision (what `H5Tset_precision` produces; the parser does not
+/// keep either field).
+fn canon_custom_float(dt: &Datatype, b: &[u8], out: &mut Vec<u8>) -> Result<(), String> {
+    let Datatype::FloatingPoint {
+        size,
+        byte_order,
+        bit_offset,
+        bit_precision,
+        exponent_location,
+        exponent_size,
+        mantissa_location,
+        mantissa_size,
+        exponent_bias,
+    } = dt
+    else {
+        unreachable!()
+    };
+    let (esize, msize) = (u32::from(*exponent_size), u32::from(*mantissa_size));
+    if esize == 0 || esize > 30 || msize > 64 {
+        return Err(format!("canon: unsupported float layout e{esize} m{msize}"));
+    }
+    let v = element_bits(b, byte_order)?;
+    let sign_pos = (u32::from(*bit_offset) + u32::from(*bit_precision)).saturating_sub(1);
+    let neg = field(v, sign_pos, 1) == 1;
+    let e = field(v, u32::from(*exponent_location), esize) as i64;
+    let m = field(v, u32::from(*mantissa_location), msize);
+    let emax = (1i64 << esize) - 1;
+    let bias = i64::from(*exponent_bias);
+    let mag = if e == emax {
+        if m == 0 { f64::INFINITY } else { f64::NAN }
+    } else if e == 0 {
+        (m as f64) * 2f64.powi((1 - bias - msize as i64) as i32)
+    } else {
+        ((1u128 << msize) as f64 + m as f64) * 2f64.powi((e - bias - msize as i64) as i32)
+    };
+    let x = if neg { -mag } else { mag };
+    match size {
+        2 => out
+            .extend_from_slice(&clawhdf5_format::float16::f32_to_f16_bits(x as f32).to_le_bytes()),
+        4 => out.extend_from_slice(&(x as f32).to_le_bytes()),
+        8 => out.extend_from_slice(&x.to_le_bytes()),
+        _ => unreachable!("ieee_layout keeps other sizes raw"),
+    }
+    Ok(())
+}
+
+/// Integers stored with a bit offset or reduced precision (N-Bit): libhdf5
+/// converts them to the full-width integer of the same size, shifting the
+/// value down and sign-extending from the top precision bit.
+fn canon_partial_int(dt: &Datatype, b: &[u8], out: &mut Vec<u8>) -> Result<(), String> {
+    let Datatype::FixedPoint {
+        size,
+        byte_order,
+        signed,
+        bit_offset,
+        bit_precision,
+    } = dt
+    else {
+        unreachable!()
+    };
+    let prec = u32::from(*bit_precision);
+    let v = element_bits(b, byte_order)?;
+    let mut x = field(v, u32::from(*bit_offset), prec);
+    if *signed && prec > 0 && prec < 128 && field(x, prec - 1, 1) == 1 {
+        x |= !0u128 << prec;
+    }
+    out.extend_from_slice(&x.to_le_bytes()[..*size as usize]);
+    Ok(())
+}
+
+fn partial_int(dt: &Datatype) -> bool {
+    matches!(dt, Datatype::FixedPoint { size, bit_offset, bit_precision, .. }
+        if *bit_offset != 0 || u32::from(*bit_precision) != size * 8)
+}
+
+fn canon_str(b: &[u8], out: &mut Vec<u8>) {
+    let cut = b.iter().position(|&c| c == 0).unwrap_or(b.len());
+    let mut s = &b[..cut];
+    while let [rest @ .., b' '] = s {
+        s = rest;
+    }
+    out.push(b'S');
+    out.extend_from_slice(&(s.len() as u32).to_le_bytes());
+    out.extend_from_slice(s);
+}
+
+fn hex(b: &[u8]) -> String {
+    b.iter().map(|x| format!("{x:02x}")).collect()
+}
+
+fn dtype_str(dt: &Datatype) -> String {
+    match dt {
+        Datatype::FixedPoint {
+            size,
+            signed,
+            byte_order,
+            ..
+        } => {
+            format!(
+                "{}{}{}",
+                bo(byte_order),
+                if *signed { "i" } else { "u" },
+                size
+            )
+        }
+        Datatype::FloatingPoint {
+            size, byte_order, ..
+        } => format!("{}f{}", bo(byte_order), size),
+        Datatype::BitField {
+            size, byte_order, ..
+        } => format!("{}b{}", bo(byte_order), size),
+        Datatype::Time { size, .. } => format!("time{size}"),
+        Datatype::String { size, .. } => format!("S{size}"),
+        Datatype::Opaque { size, .. } => format!("V{size}"),
+        Datatype::Compound { size, members } => format!(
+            "{{{}}}{size}",
+            members
+                .iter()
+                .map(|m| format!("{}:{}", m.name, dtype_str(&m.datatype)))
+                .collect::<Vec<_>>()
+                .join(",")
+        ),
+        Datatype::Reference { ref_type, .. } => format!("ref({ref_type:?})"),
+        Datatype::Enumeration { base_type, .. } => format!("enum({})", dtype_str(base_type)),
+        Datatype::VariableLength {
+            is_string: true, ..
+        } => "vlstr".into(),
+        Datatype::VariableLength { base_type, .. } => format!("vlen({})", dtype_str(base_type)),
+        Datatype::Array {
+            base_type,
+            dimensions,
+        } => format!("({}){dimensions:?}", dtype_str(base_type)),
+    }
+}
+
+fn bo(b: &DatatypeByteOrder) -> &'static str {
+    match b {
+        DatatypeByteOrder::LittleEndian => "<",
+        DatatypeByteOrder::BigEndian => ">",
+        DatatypeByteOrder::Vax => "vax",
+    }
+}
+
+fn is_group(h: &ObjectHeader) -> bool {
+    h.messages.iter().any(|m| {
+        matches!(
+            m.msg_type,
+            MessageType::LinkInfo | MessageType::Link | MessageType::SymbolTable
+        )
+    })
+}
+
+/// The probe's kind for an object header: libhdf5's object class
+/// ([`ObjectHeader::object_class`]: group, then dataset — a datatype *and* a
+/// dataspace — then named datatype), which is what h5py opens the object as.
+/// The root group, and a header with only link messages, count as groups.
+fn kind_of(h: &ObjectHeader, is_root: bool) -> &'static str {
+    match h.object_class() {
+        Some(ObjectClass::Group) => "group",
+        Some(ObjectClass::Dataset) => "dataset",
+        _ if is_root || is_group(h) => "group",
+        Some(ObjectClass::NamedDatatype) => "datatype",
+        None => "unknown",
+    }
+}
+
+fn main() {
+    install_hook();
+    let path = std::env::args().nth(1).expect("usage: probe <file>");
+    let mut top = Map::new();
+    top.insert("file".into(), Value::String(path.clone()));
+    let data = match std::fs::read(&path) {
+        Ok(d) => d,
+        Err(err) => {
+            top.insert("open_error".into(), Value::String(format!("Io({err})")));
+            println!("{}", Value::Object(top));
+            return;
+        }
+    };
+    // Every address is relative to the superblock: look at the file from
+    // there on (past any user block), as libhdf5 does.
+    let hdf5: &[u8] = match signature::find_signature(&data) {
+        Ok(off) => &data[off..],
+        Err(_) => &data,
+    };
+    let sb = guarded(|| Superblock::parse(hdf5, 0).map_err(e));
+    let sb = match sb {
+        Ok(sb) => sb,
+        Err(msg) => {
+            top.insert("open_error".into(), Value::String(msg));
+            println!("{}", Value::Object(top));
+            return;
+        }
+    };
+    // libhdf5 refuses a truncated file and reads nothing past the recorded
+    // end of file.
+    let base = (data.len() - hdf5.len()) as u64;
+    let hdf5 = match sb.data_end(base, data.len() as u64) {
+        Ok(end) => &hdf5[..end as usize],
+        Err(err) => {
+            top.insert("open_error".into(), Value::String(e(err)));
+            println!("{}", Value::Object(top));
+            return;
+        }
+    };
+    // libhdf5 decodes the superblock extension at open (an error refuses
+    // the file), and loads a metadata cache image over the file's own
+    // metadata. It loads the image only when it first reads metadata — the
+    // root group — so a file whose image it cannot load still opens and
+    // that read fails. The library decides all three cases with the same
+    // `cache_image_state`: `File` and `MmapFile` open such a file and fail
+    // every object lookup with the image's error, which is what the probe
+    // records here (on the root group, where libhdf5 reports it).
+    use clawhdf5_format::superblock_ext::{self, CacheImageState};
+    let state = match guarded(|| superblock_ext::cache_image_state(hdf5, &sb).map_err(e)) {
+        Ok(x) => x,
+        Err(msg) => {
+            top.insert("open_error".into(), Value::String(msg));
+            println!("{}", Value::Object(top));
+            return;
+        }
+    };
+    let mut image_error = None;
+    let view = match state {
+        CacheImageState::Absent => None,
+        CacheImageState::Unloadable(err) => {
+            image_error = Some(e(err));
+            None
+        }
+        CacheImageState::Loaded(image) => {
+            let mut v = hdf5.to_vec();
+            match image.block(hdf5).and_then(|b| image.apply(b, &mut v)) {
+                Ok(()) => Some(v),
+                Err(err) => {
+                    image_error = Some(e(err));
+                    None
+                }
+            }
+        }
+    };
+    let hdf5: &[u8] = view.as_deref().unwrap_or(hdf5);
+    top.insert("superblock_version".into(), json!(sb.version));
+    let ctx = Ctx {
+        data: hdf5,
+        os: sb.offset_size,
+        ls: sb.length_size,
+        base_dir: std::path::Path::new(&path)
+            .parent()
+            .map(|p| p.to_path_buf())
+            .unwrap_or_default(),
+        vl: RefCell::new(VlResolver::new(hdf5, sb.offset_size, sb.length_size)),
+    };
+    let mut objects: Vec<Value> = Vec::new();
+    let mut visited = HashSet::new();
+    let mut soft_v1 = 0u64;
+    // explicit DFS stack: (address, path)
+    let mut stack: Vec<(u64, String)> = vec![(sb.root_group_address, "/".to_string())];
+    while let Some((addr, p)) = stack.pop() {
+        if objects.len() >= MAX_OBJECTS {
+            top.insert("truncated".into(), json!(true));
+            break;
+        }
+        if !visited.insert(addr) {
+            continue;
+        }
+        let mut rec = Map::new();
+        rec.insert("path".into(), Value::String(p.clone()));
+        let r = guarded(|| {
+            if let Some(msg) = &image_error {
+                return Err(msg.clone());
+            }
+            let h = ctx.header(addr)?;
+            Ok(h)
+        });
+        let h = match r {
+            Ok(h) => h,
+            Err(msg) => {
+                rec.insert("kind".into(), Value::String("unknown".into()));
+                rec.insert("error".into(), Value::String(msg));
+                objects.push(Value::Object(rec));
+                continue;
+            }
+        };
+        let kind = kind_of(&h, addr == sb.root_group_address);
+        rec.insert("kind".into(), Value::String(kind.into()));
+        if kind == "dataset"
+            && let Err(msg) = guarded(|| ctx.read_dataset(&h, &mut rec))
+        {
+            rec.insert("error".into(), Value::String(msg));
+        }
+        // Opening a committed datatype decodes it (h5py's `f[name]` fails on
+        // one libhdf5 cannot decode), so decode it here too.
+        if kind == "datatype"
+            && let Err(msg) = guarded(|| ctx.read_named_datatype(&h))
+        {
+            rec.insert("error".into(), Value::String(msg));
+        }
+        if kind != "datatype" {
+            match guarded(|| ctx.attrs(&h)) {
+                Ok(m) => {
+                    rec.insert("attrs".into(), Value::Object(m));
+                }
+                Err(msg) => {
+                    rec.insert("attrs_error".into(), Value::String(msg));
+                }
+            }
+        }
+        if kind == "group" {
+            match guarded(|| ctx.entries(&h)) {
+                Ok(mut ents) => {
+                    ents.retain(|en| {
+                        if en.cache_type == 2 {
+                            soft_v1 += 1;
+                            false
+                        } else {
+                            true
+                        }
+                    });
+                    ents.sort_by(|a, b| a.name.cmp(&b.name));
+                    let base = if p == "/" { String::new() } else { p.clone() };
+                    for en in ents.into_iter().rev() {
+                        stack.push((en.object_header_address, format!("{base}/{}", en.name)));
+                    }
+                }
+                Err(msg) => {
+                    rec.insert("list_error".into(), Value::String(msg));
+                }
+            }
+        }
+        objects.push(Value::Object(rec));
+    }
+    if soft_v1 > 0 {
+        top.insert("v1_soft_link_entries".into(), json!(soft_v1));
+    }
+    top.insert("objects".into(), Value::Array(objects));
+    println!("{}", Value::Object(top));
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// The N-Bit float of libhdf5's `test/testfiles/le_data.h5`
+    /// (`Nbit_float_data_le`): offset 7, precision 20, sign bit 26, exponent
+    /// 20+6 (bias 31), mantissa 7+13.
+    fn nbit_f32(byte_order: DatatypeByteOrder) -> Datatype {
+        Datatype::FloatingPoint {
+            size: 4,
+            byte_order,
+            bit_offset: 7,
+            bit_precision: 20,
+            exponent_location: 20,
+            exponent_size: 6,
+            mantissa_location: 7,
+            mantissa_size: 13,
+            exponent_bias: 31,
+        }
+    }
+
+    fn canon_one(dt: &Datatype, bytes: &[u8]) -> Vec<u8> {
+        let mut out = Vec::new();
+        canon_custom_float(dt, bytes, &mut out).unwrap();
+        out
+    }
+
+    #[test]
+    fn nbit_float_canonicalises_to_the_value_libhdf5_returns() {
+        let le = nbit_f32(DatatypeByteOrder::LittleEndian);
+        let be = nbit_f32(DatatypeByteOrder::BigEndian);
+        assert!(!ieee_layout(&le));
+        // 1.0: exponent = bias, mantissa 0
+        let one: u32 = 31 << 20;
+        assert_eq!(canon_one(&le, &one.to_le_bytes()), 1.0f32.to_le_bytes());
+        assert_eq!(canon_one(&be, &one.to_be_bytes()), 1.0f32.to_le_bytes());
+        // -2.1999512 (h5py's reading of the file's -2.2): sign, e = 32, m = 819
+        let v: u32 = (1 << 26) | (32 << 20) | (819 << 7);
+        assert_eq!(
+            canon_one(&le, &v.to_le_bytes()),
+            (-2.199_951_2f32).to_le_bytes()
+        );
+        assert_eq!(canon_one(&le, &[0; 4]), 0.0f32.to_le_bytes());
+    }
+
+    #[test]
+    fn ieee_floats_keep_their_raw_bytes() {
+        let f32le = Datatype::FloatingPoint {
+            size: 4,
+            byte_order: DatatypeByteOrder::LittleEndian,
+            bit_offset: 0,
+            bit_precision: 32,
+            exponent_location: 23,
+            exponent_size: 8,
+            mantissa_location: 0,
+            mantissa_size: 23,
+            exponent_bias: 127,
+        };
+        assert!(ieee_layout(&f32le));
+    }
+
+    #[test]
+    fn kind_follows_libhdf5_object_class() {
+        use clawhdf5_format::object_header::HeaderMessage;
+        let header = |types: &[MessageType]| ObjectHeader {
+            version: 2,
+            messages: types
+                .iter()
+                .map(|&msg_type| HeaderMessage {
+                    msg_type,
+                    size: 0,
+                    flags: 0,
+                    creation_order: None,
+                    data: Vec::new(),
+                })
+                .collect(),
+            reference_count: None,
+            flags: 0,
+            access_time: None,
+            modification_time: None,
+            change_time: None,
+            birth_time: None,
+        };
+        use MessageType::*;
+        // cve-2024-33874 `/Dset1`: a datatype and a layout but no dataspace
+        // is a named datatype to libhdf5 (h5py opens it as one).
+        assert_eq!(kind_of(&header(&[Datatype, DataLayout]), false), "datatype");
+        assert_eq!(
+            kind_of(&header(&[Datatype, Dataspace, DataLayout]), false),
+            "dataset"
+        );
+        assert_eq!(kind_of(&header(&[SymbolTable]), false), "group");
+        assert_eq!(kind_of(&header(&[Link]), false), "group");
+        assert_eq!(kind_of(&header(&[]), true), "group");
+        assert_eq!(kind_of(&header(&[]), false), "unknown");
+    }
+
+    #[test]
+    fn partial_precision_int_is_shifted_and_sign_extended() {
+        let dt = Datatype::FixedPoint {
+            size: 4,
+            byte_order: DatatypeByteOrder::BigEndian,
+            signed: true,
+            bit_offset: 4,
+            bit_precision: 17,
+        };
+        assert!(partial_int(&dt));
+        let stored = (((-5i32) as u32) & 0x1_FFFF) << 4;
+        let mut out = Vec::new();
+        canon_partial_int(&dt, &stored.to_be_bytes(), &mut out).unwrap();
+        assert_eq!(out, (-5i32).to_le_bytes());
+    }
+}

conformance/ref.py

@@ -0,0 +1,325 @@
+#!/usr/bin/env python3
+"""Reference probe: same JSON as the Rust `conformance-probe`, produced with h5py.
+
+Walk: iterative DFS from '/', children in sorted (UTF-8 byte) name order, hard
+links only, each object once (first path wins, deduplicated by object identity).
+Canonical value encoding: see harness/src/main.rs.
+"""
+import hashlib
+import json
+import os
+import struct
+import sys
+
+import numpy as np
+import h5py
+
+try:
+    import hdf5plugin  # noqa: F401  registers blosc/lz4/zstd/bzip2/... filters
+except Exception:  # pragma: no cover
+    pass
+
+MAX_BYTES = 200 * 1024 * 1024
+MAX_OBJECTS = 200_000
+
+
+def canon_str(b, out):
+    if isinstance(b, str):
+        b = b.encode("utf-8", "surrogateescape")
+    b = bytes(b)
+    cut = b.find(b"\x00")
+    if cut >= 0:
+        b = b[:cut]
+    b = b.rstrip(b" ")
+    out += b"S" + struct.pack("<I", len(b)) + b
+
+
+def simple(dt):
+    if dt.fields:
+        return all(simple(dt.fields[n][0]) for n in dt.names)
+    if dt.subdtype:
+        return simple(dt.subdtype[0])
+    return dt.kind in "iufcbV"
+
+
+def packed(dt):
+    if dt.fields:
+        return np.dtype([(n, packed(dt.fields[n][0])) for n in dt.names])
+    if dt.subdtype:
+        base, shape = dt.subdtype
+        return np.dtype((packed(base), shape))
+    if dt.kind in "iufcb":
+        return dt.newbyteorder("<")
+    return dt
+
+
+# --- reference corrections ---------------------------------------------------
+# Where h5py is known to return values the file does not hold, and the right
+# values follow from what it returned, ref.py corrects them and records the
+# correction on the object ("ref_fix"), so the comparison is still a real
+# comparison and CONFORMANCE.md lists every corrected object. Each correction
+# first checks that the installed h5py still has the bug.
+
+# Corrections applied while encoding the current object.
+FIXES = set()
+_BE_VLEN_BUG = None
+
+
+def be_vlen_bug():
+    """h5py (3.16 / HDF5 2.0 at least) returns the elements of a
+    variable-length sequence whose base type is big-endian with the file's
+    big-endian bytes under a native (little-endian) dtype: a
+    `vlen_dtype('>f4')` dataset holding [1.0, 2.0] reads back as
+    [4.6e-41, 9.0e-44]. `h5dump` prints the file's values. Checked once per
+    process by writing and reading exactly that dataset in memory."""
+    global _BE_VLEN_BUG
+    if _BE_VLEN_BUG is None:
+        import io
+        try:
+            bio = io.BytesIO()
+            with h5py.File(bio, "w") as f:
+                d = f.create_dataset("v", (1,), dtype=h5py.vlen_dtype(np.dtype(">f4")))
+                d[0] = np.array([1.0, 2.0], dtype=">f4")
+            with h5py.File(bio, "r") as f:
+                got = np.asarray(f["v"][0])
+            _BE_VLEN_BUG = (got.dtype == np.dtype("<f4")
+                            and got.view(">f4").tolist() == [1.0, 2.0]
+                            and got.tolist() != [1.0, 2.0])
+        except Exception:  # noqa: BLE001
+            _BE_VLEN_BUG = False
+    return _BE_VLEN_BUG
+
+
+def unswapped(got, base):
+    """`got` is `base` (big-endian somewhere) with every field in native
+    little-endian order instead: the shape of h5py's big-endian VL bug."""
+    return base.newbyteorder("<") == got and base != got
+
+
+def canon_el(dt, val, out):
+    if dt.fields:
+        for n in dt.names:
+            canon_el(dt.fields[n][0], val[n], out)
+        return
+    if dt.subdtype:
+        base, _ = dt.subdtype
+        for x in np.asarray(val).reshape(-1):
+            canon_el(base, x, out)
+        return
+    k = dt.kind
+    if k in "iufcb":
+        out += np.asarray(val, dtype=dt).astype(dt.newbyteorder("<")).tobytes()
+    elif k == "V":
+        out += np.asarray(val, dtype=dt).tobytes()
+    elif k == "S":
+        canon_str(val, out)
+    elif k == "O":
+        if h5py.check_string_dtype(dt) is not None:
+            canon_str(val if val is not None else b"", out)
+        elif h5py.check_ref_dtype(dt) is not None:
+            out += b"R"
+        else:
+            base = h5py.check_vlen_dtype(dt)
+            if base is None:
+                raise TypeError(f"unhandled object dtype {dt!r}")
+            arr = np.asarray(val if val is not None else [])
+            if arr.dtype != base and be_vlen_bug() and unswapped(arr.dtype, base):
+                # h5py's big-endian VL bug (see be_vlen_bug): the bytes are
+                # the file's, the dtype label is wrong. Relabel, don't convert.
+                arr = arr.view(base)
+                FIXES.add("h5py-be-vlen")
+            arr = np.asarray(arr, dtype=base).reshape(-1)
+            out += b"V" + struct.pack("<I", arr.shape[0])
+            if simple(base):
+                out += arr.astype(packed(base)).tobytes()
+            else:
+                for x in arr:
+                    canon_el(base, x, out)
+    elif k == "U":
+        canon_str(str(val), out)
+    else:
+        raise TypeError(f"unhandled dtype kind {k} ({dt!r})")
+
+
+def has_obj(dt):
+    if dt.fields:
+        return any(has_obj(dt.fields[n][0]) for n in dt.names)
+    if dt.subdtype:
+        return has_obj(dt.subdtype[0])
+    return dt.kind == "O"
+
+
+def note_conversion(tid, dt, rec):
+    """h5py converts some file types (FP8, bfloat16, x87 long double, ...) to a
+    different-sized numpy type; then value bytes are not comparable."""
+    try:
+        if not has_obj(dt) and tid.get_size() != dt.itemsize:
+            rec["converted"] = f"file type size {tid.get_size()} -> numpy {dt} ({dt.itemsize})"
+    except Exception:  # noqa: BLE001
+        pass
+
+
+def hash_values(arr, dt, rec):
+    # h5py expands an HDF5 array element type into trailing array dims, a
+    # nested array type (an array of arrays) into all of them. Converting the
+    # expanded array back to the inner subarray type would broadcast every
+    # element into a whole subarray, so strip every level.
+    while dt.subdtype is not None:
+        dt = dt.subdtype[0]
+    arr = np.asarray(arr, dtype=dt)
+    FIXES.clear()
+    if simple(dt):
+        c = np.ascontiguousarray(arr).astype(packed(dt)).tobytes()
+    else:
+        out = bytearray()
+        for x in arr.reshape(-1):
+            canon_el(dt, x, out)
+        c = bytes(out)
+    if FIXES:
+        rec["ref_fix"] = sorted(FIXES)
+    rec["hash"] = hashlib.sha256(c).hexdigest()
+    rec["head"] = c[:48].hex()
+
+
+def err(e):
+    s = f"{type(e).__name__}: {e}"
+    return s.splitlines()[0][:400] if s else type(e).__name__
+
+
+def shape_of(s):
+    return "null" if s is None else list(s)
+
+
+def n_bytes(shape, tid):
+    n = 1
+    for d in shape or ():
+        n *= d
+    return n * tid.get_size()
+
+
+def read_attrs(obj):
+    out = {}
+    names = sorted(obj.attrs.keys(), key=lambda s: s.encode("utf-8", "surrogateescape"))
+    for name in names:
+        rec = {}
+        try:
+            aid = obj.attrs.get_id(name)
+            rec["dtype"] = str(aid.dtype)
+            rec["shape"] = shape_of(aid.shape)
+            note_conversion(aid.get_type(), aid.dtype, rec)
+            if aid.shape is None:
+                hash_values(np.empty((0,), dtype=aid.dtype), aid.dtype, rec)
+            else:
+                val = obj.attrs[name]
+                hash_values(val, aid.dtype, rec)
+        except Exception as e:  # noqa: BLE001
+            rec = {"error": err(e)}
+        out[name] = rec
+    return out
+
+
+def main(path):
+    top = {"file": path}
+    try:
+        f = h5py.File(path, "r")
+    except Exception as e:  # noqa: BLE001
+        top["open_error"] = err(e)
+        print(json.dumps(top))
+        return
+    objects = []
+    seen = set()
+    # Objects h5py cannot open have no ObjectID to deduplicate by; they are
+    # deduplicated by the address their hard link points at instead, as the
+    # probe deduplicates every object by header address.
+    seen_unopenable = set()
+    stack = [("/", None, None)]
+    while stack:
+        p, obj, link_addr = stack.pop()
+        if len(objects) >= MAX_OBJECTS:
+            top["truncated"] = True
+            break
+        rec = {"path": p}
+        try:
+            if obj is None:
+                obj = f[p]
+            key = hash(obj.id)  # h5py ObjectID hash = (fileno, object address/token)
+        except Exception as e:  # noqa: BLE001
+            if link_addr is not None:
+                if link_addr in seen_unopenable:
+                    continue
+                seen_unopenable.add(link_addr)
+            rec["kind"] = "unknown"
+            rec["error"] = err(e)
+            objects.append(rec)
+            continue
+        if key in seen:
+            continue
+        seen.add(key)
+        if isinstance(obj, h5py.Dataset):
+            kind = "dataset"
+        elif isinstance(obj, h5py.Group):
+            kind = "group"
+        elif isinstance(obj, h5py.Datatype):
+            kind = "datatype"
+        else:
+            kind = "unknown"
+        rec["kind"] = kind
+        if kind == "dataset":
+            try:
+                dt = obj.dtype
+                rec["dtype"] = str(dt)
+                rec["shape"] = shape_of(obj.shape)
+                note_conversion(obj.id.get_type(), dt, rec)
+                if obj.shape is None:
+                    hash_values(np.empty((0,), dtype=dt), dt, rec)
+                elif n_bytes(obj.shape, obj.id.get_type()) > MAX_BYTES:
+                    rec["skipped"] = "too large"
+                else:
+                    arr = np.empty(obj.shape, dtype=dt)
+                    if arr.size:
+                        try:
+                            obj.read_direct(arr)
+                        except Exception:  # noqa: BLE001
+                            arr = obj[()]
+                    hash_values(arr, dt, rec)
+            except Exception as e:  # noqa: BLE001
+                rec["error"] = err(e)
+        if kind != "datatype":
+            try:
+                rec["attrs"] = read_attrs(obj)
+            except Exception as e:  # noqa: BLE001
+                rec["attrs_error"] = err(e)
+        if kind == "group":
+            try:
+                names = sorted(obj.keys(), key=lambda s: s.encode("utf-8", "surrogateescape"))
+                base = "" if p == "/" else p
+                kids = []
+                for n in names:
+                    # The link's own type: `obj.get(n, getlink=True)` reports
+                    # a user-defined link (type 64-255) as a HardLink.
+                    try:
+                        info = obj.id.links.get_info(n.encode("utf-8", "surrogateescape"))
+                    except Exception:  # noqa: BLE001
+                        info = None
+                    if info is not None and info.type != h5py.h5l.TYPE_HARD:
+                        continue
+                    addr = info.u if info is not None else None
+                    kids.append((f"{base}/{n}", addr))
+                for k, addr in reversed(kids):
+                    stack.append((k, None, addr))
+            except Exception as e:  # noqa: BLE001
+                rec["list_error"] = err(e)
+        objects.append(rec)
+    top["objects"] = objects
+    print(json.dumps(top), flush=True)
+    # Exit without tearing down the h5py objects: freeing them for some files
+    # that hold references (hdf5's h5repack_attr_refs.h5, cve-2024-32623.h5)
+    # makes libhdf5 2.0 abort with "free(): chunks in smallbin corrupted"
+    # about half the time. That happens after the reading is done, so it says
+    # nothing about what h5py read, but it flipped those files between ok and
+    # h5py-cannot-read from one run to the next.
+    os._exit(0)
+
+
+if __name__ == "__main__":
+    main(sys.argv[1])

conformance/ref_bugs.py

@@ -0,0 +1,105 @@
+#!/usr/bin/env python3
+"""ref_bugs.py <corpus_dir>: re-check the objects h5py reads only through a
+libhdf5 bug.
+
+For each object of READ_BUGS (below), h5py reads it in several fresh
+processes whose heaps differ: h5py imported before numpy (three runs, plus
+two with glibc's MALLOC_PERTURB_, which fills newly allocated and freed heap
+blocks with a byte pattern) and numpy imported first. Values the file
+determines come out the same every time. An object whose values differ
+between those runs is read from memory the file does not determine — an
+over-read or an uninitialised buffer in libhdf5 — so the values h5py reports
+for it are not the file's, and clawhdf5 refusing the object is not a
+clawhdf5 error. compare.py classifies a file as `ref-bug` only on objects
+confirmed that way in the same run (`$OUT/ref_bugs.json`); an object whose
+reading turns out stable stays an our-error.
+
+Run by conformance/run.sh; on its own it is the reproducer (JSON on stdout).
+"""
+import concurrent.futures
+import json
+import os
+import subprocess
+import sys
+
+# (file, object) -> what goes wrong. Checked 2026-09-27 against HDF5 2.0.0
+# (h5py 3.16), h5dump 1.14.6 and the HDFGroup/hdf5 sources (tag hdf5_1_14_6
+# and develop); see docs/known-issues.md, "Conformance: the last non-ok files".
+READ_BUGS = {
+    ("cve_hdf5/cvefiles/cve-2025-2308.h5", "/Scale_offset_long_long_data_le"):
+        "the first chunk records minbits 11: its 12 values need 17 bytes of codes, and the "
+        "26-byte chunk holds 5 after its 21-byte header; libhdf5's scale-offset decoder reads "
+        "past its buffer, and develop refuses the chunk (\"Buffer too short\")",
+    ("cve_hdf5/cvefiles/cve-2025-44904.h5", "/Scale_offset_float_data_le"):
+        "unfiltered chunks stored as 38 and 37 bytes for 48-byte chunks: 1.14/2.0 read the "
+        "stored bytes into a buffer of that size and use it as the whole chunk "
+        "(H5D__chunk_lock), so the rest is heap memory; develop refuses them (\"incorrect chunk "
+        "size returned from index for unfiltered chunk\")",
+    ("hdf5/test/testfiles/bad_nbit_parms_walk.h5", "/Nbit_int_data_le"):
+        "the N-Bit parameter list holds 7 values (cd_values[0] = 7) where an integer needs 8: "
+        "the decoder takes the bit offset from cd_values[7], past the list; libhdf5's own test "
+        "(`test_filter_bad_params`, test/dsets.c on develop) requires the read to fail",
+}
+
+# (which module is imported first, MALLOC_PERTURB_)
+RUNS = [("h5py", None), ("h5py", None), ("h5py", None), ("h5py", "170"), ("h5py", "255"),
+        ("numpy", None)]
+
+READ = r"""
+import hashlib, sys
+if sys.argv[3] == "h5py":
+    import h5py, numpy as np
+else:
+    import numpy as np, h5py
+try:
+    import hdf5plugin  # noqa: F401
+except Exception:
+    pass
+try:
+    with h5py.File(sys.argv[1], "r") as f:
+        a = np.ascontiguousarray(f[sys.argv[2]][()])
+    print("values " + hashlib.sha256(a.tobytes()).hexdigest()[:16])
+except Exception as e:
+    print("error " + (str(e).splitlines() or [type(e).__name__])[0][:120])
+"""
+
+
+def read_once(path, obj, first, perturb):
+    env = dict(os.environ)
+    env.pop("MALLOC_PERTURB_", None)
+    if perturb:
+        env["MALLOC_PERTURB_"] = perturb
+    try:
+        p = subprocess.run([sys.executable, "-c", READ, path, obj, first], env=env,
+                           capture_output=True, text=True, timeout=60)
+        out = p.stdout.strip().splitlines()
+        return out[-1] if out else f"exit {p.returncode}"
+    except subprocess.TimeoutExpired:
+        return "timeout"
+
+
+def check(corpus, key):
+    f, obj = key
+    path = os.path.join(corpus, f)
+    rec = {"file": f, "object": obj, "why": READ_BUGS[key]}
+    if not os.path.exists(path):
+        return rec | {"missing": True, "confirmed": False}
+    runs = [{"first": a, "malloc_perturb": p, "outcome": read_once(path, obj, a, p)} for a, p in RUNS]
+    distinct = sorted({r["outcome"] for r in runs})
+    return rec | {
+        "runs": runs,
+        "distinct": len(distinct),
+        "confirmed": len(distinct) > 1 and any(o.startswith("values ") for o in distinct),
+    }
+
+
+def main():
+    corpus = sys.argv[1]
+    keys = list(READ_BUGS)
+    with concurrent.futures.ThreadPoolExecutor(max_workers=len(keys)) as ex:
+        out = list(ex.map(lambda k: check(corpus, k), keys))
+    print(json.dumps({"read_bugs": out}, indent=1))
+
+
+if __name__ == "__main__":
+    main()

conformance/report.py

@@ -0,0 +1,361 @@
+#!/usr/bin/env python3
+"""report.py <results_dir> <CONFORMANCE.md> <corpus_dir>
+
+Render the sweep's results (compare.py's results.json plus the raw per-side
+runs) as CONFORMANCE.md, and write <results_dir>/report-meta.json (commit,
+date, versions) for check.py --update.
+"""
+import collections
+import datetime
+import json
+import os
+import platform
+
+import subprocess
+import sys
+
+import h5py
+import numpy
+
+try:
+    import hdf5plugin
+    HDF5PLUGIN = hdf5plugin.version
+except Exception:  # noqa: BLE001
+    HDF5PLUGIN = "not installed"
+
+R, OUT_MD, CORPUS = sys.argv[1], sys.argv[2], sys.argv[3]
+HERE = os.path.dirname(os.path.abspath(__file__))
+ROOT = os.path.dirname(HERE)
+CLASSES = ["ok", "our-error", "mismatch", "h5py-cannot-read", "ref-bug", "panic", "hang", "crash", "oom"]
+
+
+def sh(*cmd, cwd=ROOT):
+    try:
+        return subprocess.run(cmd, cwd=cwd, capture_output=True, text=True, timeout=30).stdout.strip()
+    except Exception:  # noqa: BLE001
+        return ""
+
+
+def cpu_model():
+    try:
+        for ln in open("/proc/cpuinfo"):
+            if ln.startswith(("model name", "Model")):
+                return ln.split(":", 1)[1].strip()
+    except OSError:
+        pass
+    return platform.processor() or "unknown"
+
+
+def mem_gib():
+    try:
+        for ln in open("/proc/meminfo"):
+            if ln.startswith("MemTotal:"):
+                return f"{int(ln.split()[1]) / 1048576:.0f} GiB"
+    except OSError:
+        pass
+    return "?"
+
+
+res = json.load(open(os.path.join(R, "results.json")))
+meta_run = json.load(open(os.path.join(R, "meta.json"))) if os.path.exists(os.path.join(R, "meta.json")) else {}
+rows = res["rows"]
+issues = res.get("issues", {})
+
+# safe.directory: a checkout owned by another user (a container) is still ours to read
+commit = sh("git", "-c", "safe.directory=*", "rev-parse", "HEAD") or os.environ.get("GITHUB_SHA", "unknown")
+lib_dirty = sh("git", "-c", "safe.directory=*", "status", "--porcelain", "--", "crates", "Cargo.toml")
+h5dump_v = sh("h5dump", "--version").replace("h5dump: ", "")
+meta = {
+    "date": datetime.datetime.now(datetime.timezone.utc).strftime("%Y-%m-%d %H:%M UTC"),
+    "commit": commit + (" (library sources modified)" if lib_dirty else ""),
+    "reference": f"h5py {h5py.__version__} / HDF5 {h5py.version.hdf5_version}",
+}
+json.dump(meta, open(os.path.join(R, "report-meta.json"), "w"), indent=1)
+
+pins = []
+for ln in open(os.path.join(HERE, "corpus.txt")):
+    if ln.strip() and not ln.lstrip().startswith("#"):
+        name, url, rev, root, *_ = ln.split()
+        pins.append((name, url, rev, root))
+
+by_corpus = collections.defaultdict(collections.Counter)
+for r in rows:
+    by_corpus[r["corpus"]][r["class"]] += 1
+total = collections.Counter(r["class"] for r in rows)
+
+
+def ex_list(files, n=3):
+    s = ", ".join(f"`{f}`" for f in files[:n])
+    return s + (f" (+{len(files) - n} more)" if len(files) > n else "")
+
+
+# --- reference bugs ---------------------------------------------------------
+# ref_bugs.py's re-check of the objects h5py reads only through a libhdf5 bug
+# (compare.py classifies on the confirmed ones), and the objects whose h5py
+# values ref.py corrected (compare.py's ref_fixes).
+try:
+    ref_bugs = json.load(open(os.path.join(R, "ref_bugs.json")))["read_bugs"]
+except (OSError, ValueError, KeyError):
+    ref_bugs = []
+ref_fixes = res.get("ref_fixes", [])
+
+
+# --- the CVE corpus: clawhdf5 vs h5dump vs h5py ------------------------------
+def side(run, name):
+    p = os.path.join(R, "runs", run, name)
+    if not os.path.exists(p + ".rc"):
+        return None
+    rc = int(open(p + ".rc").read().strip() or -1)
+    err = open(p + ".err", errors="replace").read()
+    try:
+        j = json.load(open(p + ".json"))
+    except Exception:  # noqa: BLE001
+        j = None
+    return rc, err, j
+
+
+def outcome(s, rust=False):
+    """-> (bucket, text). bucket in read / error / panic / crash / hang / oom."""
+    if s is None:
+        return "missing", "not run"
+    rc, err, j = s
+    if rc in (137, 124):
+        return "hang", "hang (killed at timeout)"
+    if "memory allocation of" in err or "MemoryError" in err or "bad_alloc" in err or "Cannot allocate" in err:
+        return "oom", "out of memory"
+    if rust and (rc == 101 or "PANIC:" in err):
+        return "panic", "panic"
+    if "overflowed its stack" in err:
+        return "crash", "stack overflow"
+    if rc == 139:
+        return "crash", "SIGSEGV"
+    if rc == 134:
+        return "crash", "SIGABRT" + (" (heap corruption)" if ("corrupted" in err or "free()" in err) else "")
+    if rc > 128:
+        return "crash", f"signal {rc - 128}"
+    if j is None:
+        return ("error", "error exit") if rc in (0, 1) else ("crash", f"exit {rc}")
+    if "open_error" in j:
+        return "error", "open error"
+    objs = j.get("objects", [])
+    ne = sum(1 for o in objs for k in ("error", "attrs_error", "list_error") if k in o)
+    ne += sum(1 for o in objs for a in (o.get("attrs") or {}).values() if "error" in a)
+    return "read", f"read {len(objs)} obj" + (f", {ne} errors" if ne else "")
+
+
+def h5dump_outcome(s):
+    if s is None:
+        return "missing", "not run"
+    rc, err, _ = s
+    if rc in (137, 124):
+        return "hang", "hang (killed at timeout)"
+    if "memory allocation" in err or "Cannot allocate" in err:
+        return "oom", "out of memory"
+    if rc == 139:
+        return "crash", "SIGSEGV"
+    if rc == 134:
+        return "crash", "SIGABRT" + (" (heap corruption)" if ("corrupted" in err or "free()" in err) else "")
+    if rc > 128:
+        return "crash", f"signal {rc - 128}"
+    return ("read", "ok") if rc == 0 else ("error", "error exit")
+
+
+cve_rows = []
+buckets = {"clawhdf5": collections.Counter(), "h5dump": collections.Counter(), "h5py": collections.Counter()}
+ours_panic = {r["file"] for r in rows if r["class"] == "panic"}
+for r in rows:
+    if r["corpus"] != "cve_hdf5":
+        continue
+    run = r["file"].replace("/", "__")
+    o = outcome(side(run, "ours"), rust=True)
+    if o[0] == "read" and r["file"] in ours_panic:
+        o = ("panic", "caught panic")
+    p = outcome(side(run, "ref"))
+    d = h5dump_outcome(side(run, "h5dump"))
+    buckets["clawhdf5"][o[0]] += 1
+    buckets["h5py"][p[0]] += 1
+    buckets["h5dump"][d[0]] += 1
+    cve_rows.append((r["file"].split("/", 1)[1], d[1], p[1], o[1], r["class"]))
+
+# --- render -----------------------------------------------------------------
+L = []
+w = L.append
+w("# clawhdf5 conformance report")
+w("")
+w("Every HDF5 file of eight public corpora (pinned by commit) is read twice — by")
+w("clawhdf5 (`conformance/probe`, the same `clawhdf5-format` calls the facade")
+w("makes) and by h5py/libhdf5 (`conformance/ref.py`) — and the two readings are")
+w("compared object by object: the set of hard-linked objects, each dataset's and")
+w("attribute's shape, and a SHA-256 of its values in a canonical encoding. The")
+w("CVE corpus is also run through `h5dump`. Each side runs under a timeout and an")
+w("address-space limit, so a hang, crash or runaway allocation is recorded, not")
+w("fatal. This file is generated by `conformance/run.sh`; do not edit it by hand.")
+w("")
+w("## Run")
+w("")
+w("| | |")
+w("|---|---|")
+w(f"| date | {meta['date']} |")
+w(f"| clawhdf5 commit | `{meta['commit']}` |")
+w(f"| machine | `{platform.node()}`: {cpu_model()}, {os.cpu_count()} CPUs, {mem_gib()}, {platform.system()} {platform.release()} {platform.machine()} |")
+w(f"| command | `{os.environ.get('CONFORMANCE_CMD', 'conformance/run.sh')}` |")
+w(f"| rustc | {sh('rustc', '-V')} |")
+w(f"| reference | h5py {h5py.__version__}, HDF5 {h5py.version.hdf5_version}, numpy {numpy.__version__}, hdf5plugin {HDF5PLUGIN}, Python {platform.python_version()} |")
+w(f"| h5dump | {h5dump_v} (CVE corpus only) |")
+if meta_run:
+    w(f"| limits | {meta_run.get('timeout_s')} s timeout (SIGKILL), {int(meta_run.get('mem_kb', 0)) // 1024} MiB address space, per process; {meta_run.get('jobs')} files in parallel |")
+    w(f"| runtime | {meta_run.get('probe_seconds')} s probing + comparing ({meta_run.get('build_seconds')} s fetch/build before it) |")
+w("")
+w("## Results")
+w("")
+w("A file's class is the first that applies:")
+w("")
+w("- **panic / hang / crash / oom** — clawhdf5 panicked (caught per object or not), hit the timeout, died on a signal, or failed an allocation. The CI gate fails on any of these.")
+w("- **h5py-cannot-read** — libhdf5 could not open the file (or itself crashed or hung). Nothing to compare against; most are the deliberately malformed CVE reproducers.")
+w("- **ref-bug** — every difference is an object clawhdf5 refuses that h5py reads only through a libhdf5 bug: the values h5py returns for it change with the reading process's heap, re-checked in every run (see *Reference bugs*).")
+w("- **our-error** — clawhdf5 returned an error for something h5py reads.")
+w("- **mismatch** — both read it, but the shapes, values, object set or attribute set differ.")
+w("- **ok** — every object h5py reads, clawhdf5 reads identically.")
+w("")
+w("| corpus | files | " + " | ".join(CLASSES) + " |")
+w("|---" * (len(CLASSES) + 2) + "|")
+for c in sorted(by_corpus):
+    cnt = by_corpus[c]
+    w(f"| {c} | {sum(cnt.values())} | " + " | ".join(str(cnt.get(k, 0)) for k in CLASSES) + " |")
+w(f"| **all** | **{len(rows)}** | " + " | ".join(f"**{total.get(k, 0)}**" for k in CLASSES) + " |")
+w("")
+nonok = total.get("our-error", 0) + total.get("mismatch", 0)
+w(f"**Our errors and mismatches: {nonok}.** Files not ok: "
+  + (", ".join(f"{total[c]} {c}" for c in CLASSES if c != "ok" and total.get(c)) or "none") + "."
+  + (f" {len(ref_fixes)} object(s) were compared against h5py's values corrected for a known h5py bug"
+     f" ({sum(1 for x in ref_fixes if x[3])} identical to clawhdf5's; see *Reference bugs*)." if ref_fixes else ""))
+w("")
+w("Corpora (fetched by `conformance/fetch-corpus.sh` into the gitignored `conformance/.cache/`):")
+w("")
+w("| corpus | source | commit |")
+w("|---|---|---|")
+for name, url, rev, root in pins:
+    w(f"| {name} | {url.removesuffix('.git')}" + ("" if root == "." else f" (`{root}`)") + f" | `{rev[:12]}` |")
+w("")
+
+w("## Panics, hangs, crashes, out-of-memory")
+w("")
+if not res["panics"]:
+    w("None.")
+else:
+    for p in res["panics"]:
+        w(f"- `{p['file']}` [{p['class']}] {p['detail']}")
+w("")
+
+w("## Our-error root causes")
+w("")
+if res["root_causes"]:
+    w("Grouped by normalised error message. *files* counts files whose class this cause affects.")
+    w("")
+    w("| files | objects | error | examples |")
+    w("|---:|---:|---|---|")
+    for k, v in res["root_causes"].items():
+        w(f"| {v['files']} | {v['count']} | `{k.replace('|', '/')}` | {ex_list(v['file_list'])} |")
+else:
+    w("None.")
+w("")
+w("## Mismatch root causes")
+w("")
+if res["mismatch_causes"]:
+    w("| files | objects | cause | examples |")
+    w("|---:|---:|---|---|")
+    for k, v in res["mismatch_causes"].items():
+        w(f"| {v['files']} | {v['count']} | `{k.replace('|', '/')}` | {ex_list(v['file_list'])} |")
+else:
+    w("None.")
+w("")
+
+w("## CVE corpus: clawhdf5 vs h5dump vs h5py")
+w("")
+w(f"The {len(cve_rows)} files of [HDFGroup/cve_hdf5](https://github.com/HDFGroup/cve_hdf5) — reproducers for")
+w("published libhdf5 CVEs and fuzzer finds. *read* = produced output (possibly with per-object")
+w("errors), *error* = refused cleanly. h5dump exits non-zero on any error anywhere in a file, so")
+w("its read/error split is not comparable with the other two rows; the panic, crash, hang and oom")
+w("columns are.")
+w("")
+w("| tool | read | error | panic | crash | hang | oom |")
+w("|---|---:|---:|---:|---:|---:|---:|")
+for tool, label in (("clawhdf5", "clawhdf5"), ("h5dump", f"h5dump {h5dump_v.split()[-1] if h5dump_v else ''}"),
+                    ("h5py", f"h5py {h5py.__version__} / HDF5 {h5py.version.hdf5_version}")):
+    b = buckets[tool]
+    w(f"| {label} | " + " | ".join(str(b.get(k, 0)) for k in ("read", "error", "panic", "crash", "hang", "oom")) + " |")
+w("")
+w("<details><summary>Per-file outcomes</summary>")
+w("")
+w("| file | h5dump | h5py | clawhdf5 | class |")
+w("|---|---|---|---|---|")
+for f, d, p, o, cls in cve_rows:
+    w(f"| {f} | {d} | {p} | {o} | {cls} |")
+w("")
+w("</details>")
+w("")
+
+w("## Reference bugs")
+w("")
+w("### Objects h5py reads only through a libhdf5 bug (*ref-bug*)")
+w("")
+w("clawhdf5 refuses these objects; h5py 3.16 / HDF5 2.0 returns values for them. `conformance/ref_bugs.py`")
+w("re-reads each with h5py in six fresh processes whose heaps differ (h5py imported before numpy, three")
+w("times and twice more with `MALLOC_PERTURB_`, and numpy imported first). Values the file determines")
+w("come out the same every time; these do not, so they are memory libhdf5 over-reads, not the file's")
+w("data. A file is *ref-bug* only while every one of its differences is such an object confirmed in")
+w("the same run; an object that reads the same every time goes back to *our-error*. Reproducer:")
+w("`python conformance/ref_bugs.py conformance/.cache/corpus` (prints every read's outcome).")
+w("")
+w("| file | object | distinct results in 6 reads | confirmed | what goes wrong |")
+w("|---|---|---:|---|---|")
+for b in ref_bugs:
+    n = "missing" if b.get("missing") else b.get("distinct", "?")
+    w(f"| `{b['file']}` | `{b['object']}` | {n} | {'yes' if b.get('confirmed') else '**no**'} | {b['why']} |")
+w("")
+w("### Values corrected for a known h5py bug")
+w("")
+w("- **h5py big-endian variable-length sequences.** h5py returns the elements of a VL sequence")
+w("  whose base type is big-endian with the file's big-endian bytes but a native (little-endian)")
+w("  numpy dtype: a `h5py.vlen_dtype(np.dtype('>f4'))` dataset holding `[1.0, 2.0]` reads back as")
+w("  `[4.6e-41, 9.0e-44]`; `h5dump` prints the file's values. `ref.py` checks that the installed")
+w("  h5py still does this (by writing and reading exactly that dataset in memory) and, if so,")
+w("  relabels such elements with the file's byte order before hashing, so the values are still")
+w("  compared. Corrected objects: "
+  + (", ".join(f"`{f}` `{p}` ({'same as clawhdf5' if same else '**differs from clawhdf5**'})"
+               for f, p, _, same in ref_fixes) if ref_fixes else "none") + ".")
+w("")
+w("## Other comparison rules")
+w("")
+w("- **Non-IEEE floats and partial-precision integers (N-Bit).** libhdf5 converts a float whose")
+w("  bit layout is not IEEE (e.g. `H5Tset_precision` for the N-Bit filter) or an integer with a")
+w("  bit offset / reduced precision into the plain numpy type of the same size. The probe")
+w("  compares such values as converted numbers, not raw file bytes (before 2026-09-25 it compared")
+w("  raw bytes, which reported every N-Bit float dataset as a mismatch).")
+if res["incomparable"]:
+    w("- **Types h5py widens.** Where h5py reads a type into a numpy type of a different size")
+    w("  (FP8 -> float16, bfloat16 -> float32, x87 long double -> float128) the values are not")
+    w("  compared (shape and presence still are): "
+      + ", ".join(f"{k} ({n}x)" for k, n in res["incomparable"]) + ".")
+w("- **References** are compared by presence only (`R`), not by target.")
+w("")
+if res.get("ref_only_errors"):
+    w("## Objects h5py fails on but clawhdf5 reads")
+    w("")
+    for k, n in res["ref_only_errors"][:15]:
+        w(f"- {n} x `{k}`")
+    w("")
+w("## Reproduce")
+w("")
+w("```sh")
+w("# needs: Rust, python3 with h5py numpy hdf5plugin (conformance/requirements.txt), h5dump (hdf5-tools), git")
+w("CLAWHDF5_PYTHON=/path/to/venv/bin/python conformance/run.sh")
+w("```")
+w("")
+w("The corpus (about 450 MB of sparse checkouts) is cached in `conformance/.cache/`; results for")
+w("every file, both sides' raw JSON and stderr, are in `conformance/.cache/results/`.")
+w("`conformance/baseline.json` holds the ok files the nightly CI job (`.gitea/workflows/conformance.yml`)")
+w("must keep; `conformance/run.sh --update-baseline` rewrites it.")
+
+with open(OUT_MD, "w") as fh:
+    fh.write("\n".join(L) + "\n")

conformance/requirements.txt

@@ -0,0 +1,6 @@
+# The reference side of the conformance sweep. Pinned so the nightly job and a
+# local run compare against the same libhdf5 (h5py wheels bundle it).
+h5py==3.16.0
+numpy==2.5.3
+hdf5plugin==7.1.0
+netCDF4==1.7.4

conformance/run.sh

@@ -0,0 +1,90 @@
+#!/usr/bin/env bash
+# conformance/run.sh — the clawhdf5 conformance sweep, end to end.
+#
+#   fetch the pinned corpora (cached) -> build the probe -> probe every file
+#   with clawhdf5 and with h5py (and h5dump for the CVE corpus), each under a
+#   timeout and a memory limit -> compare -> write CONFORMANCE.md -> check the
+#   result against conformance/baseline.json.
+#
+# Usage: conformance/run.sh [--no-fetch] [--no-report] [--update-baseline]
+#
+# Environment:
+#   CLAWHDF5_PYTHON    python with h5py, numpy, hdf5plugin (default: repo .venv, then python3)
+#   CONFORMANCE_CACHE  corpus / build / results cache (default: conformance/.cache)
+#   CONFORMANCE_OUT    results directory (default: $CONFORMANCE_CACHE/results)
+#   CONFORMANCE_REPORT report path (default: CONFORMANCE.md at the repo root)
+#   JOBS               parallel files (default: nproc)
+#   CONFORMANCE_PROBE  use this prebuilt probe binary instead of building one
+#   TMO / MEM_KB       per-process timeout in seconds (20) / address-space limit in KiB (4 GiB)
+#
+# Exit status: 0 = gate passed; 1 = a panic/hang/crash/oom in clawhdf5, or the
+# ok count fell below the baseline, or a baseline-ok file regressed; 2 = setup error.
+set -euo pipefail
+HERE="$(cd "$(dirname "$0")" && pwd)"
+ROOT="$(cd "$HERE/.." && pwd)"
+FETCH=1 REPORT=1 UPDATE=0
+for a in "$@"; do
+  case "$a" in
+    --no-fetch) FETCH=0 ;;
+    --no-report) REPORT=0 ;;
+    --update-baseline) UPDATE=1 ;;
+    -h|--help) sed -n '2,23p' "$0"; exit 0 ;;
+    *) echo "unknown argument: $a" >&2; exit 2 ;;
+  esac
+done
+
+export PATH="$HOME/.cargo/bin:$PATH"
+CACHE="${CONFORMANCE_CACHE:-$HERE/.cache}"
+mkdir -p "$CACHE"; CACHE="$(cd "$CACHE" && pwd)"
+OUT="${CONFORMANCE_OUT:-$CACHE/results}"
+REPORT_PATH="${CONFORMANCE_REPORT:-$ROOT/CONFORMANCE.md}"
+JOBS="${JOBS:-$(nproc 2>/dev/null || echo 4)}"
+if [ -n "${CLAWHDF5_PYTHON:-}" ]; then PY="$CLAWHDF5_PYTHON"
+elif [ -x "$ROOT/.venv/bin/python" ]; then PY="$ROOT/.venv/bin/python"
+else PY="$(command -v python3)"; fi
+export PY TMO="${TMO:-20}" MEM_KB="${MEM_KB:-4194304}"
+command -v h5dump >/dev/null || { echo "error: h5dump not found (install hdf5-tools)" >&2; exit 2; }
+"$PY" -c 'import h5py, numpy, hdf5plugin' || { echo "error: $PY lacks h5py/numpy/hdf5plugin" >&2; exit 2; }
+
+t0=$(date +%s)
+[ "$FETCH" = 1 ] && bash "$HERE/fetch-corpus.sh" "$CACHE"
+C="$CACHE/corpus"
+[ -d "$C" ] || { echo "error: no corpus in $C (run without --no-fetch)" >&2; exit 2; }
+
+if [ -n "${CONFORMANCE_PROBE:-}" ]; then
+  export PROBE="$CONFORMANCE_PROBE"   # a prebuilt probe, e.g. an older one for a before/after
+else
+  echo "== building the probe"
+  CARGO_TARGET_DIR="${CARGO_TARGET_DIR:-$CACHE/target}" \
+    cargo build -q --release --manifest-path "$HERE/probe/Cargo.toml"
+  export PROBE="${CARGO_TARGET_DIR:-$CACHE/target}/release/conformance-probe"
+fi
+t1=$(date +%s)
+
+rm -rf "$OUT"; mkdir -p "$OUT"
+"$PY" "$HERE/list_files.py" "$C" > "$OUT/files.txt"
+echo "== probing $(wc -l <"$OUT/files.txt") files, $JOBS at a time (timeout ${TMO}s, limit $((MEM_KB / 1024)) MiB)"
+export C OUT HERE
+# The shell's "Segmentation fault (core dumped)" notices go to probe.log; the
+# signals themselves are recorded in each side's .rc.
+xargs -a "$OUT/files.txt" -d '\n' -P "$JOBS" -I{} bash -c '
+  f="$1"; d="$OUT/runs/${f//\//__}"
+  case "$f" in cve_hdf5/*) export WITH_H5DUMP=1 ;; esac
+  "$HERE/run_one.sh" "$C/$f" "$d"' _ {} 2>"$OUT/probe.log"
+echo "== re-checking the objects h5py reads only through a libhdf5 bug"
+"$PY" "$HERE/ref_bugs.py" "$C" > "$OUT/ref_bugs.json" 2> "$OUT/ref_bugs.err" || true
+echo "== comparing"
+"$PY" "$HERE/compare.py" "$OUT" >/dev/null
+t2=$(date +%s)
+cat > "$OUT/meta.json" <<EOF
+{"build_seconds": $((t1 - t0)), "probe_seconds": $((t2 - t1)), "jobs": $JOBS, "timeout_s": $TMO, "mem_kb": $MEM_KB}
+EOF
+export CONFORMANCE_CMD="${CONFORMANCE_CMD:-conformance/run.sh${*:+ $*}}"
+if [ "$REPORT" = 1 ]; then
+  "$PY" "$HERE/report.py" "$OUT" "$REPORT_PATH" "$C"
+  echo "== wrote $REPORT_PATH"
+fi
+if [ "$UPDATE" = 1 ]; then
+  "$PY" "$HERE/check.py" "$OUT" "$HERE/baseline.json" --update
+fi
+"$PY" "$HERE/check.py" "$OUT" "$HERE/baseline.json"

conformance/run_one.sh

@@ -0,0 +1,27 @@
+#!/usr/bin/env bash
+# run_one.sh <file> <outdir>
+#
+# Probe one file with clawhdf5 (PROBE) and with h5py (PY ref.py), and with
+# h5dump too when WITH_H5DUMP is set. Each side runs under a timeout (TMO
+# seconds, SIGKILL) and an address-space limit (MEM_KB), with core dumps off.
+# Writes <outdir>/<side>.{json,err,rc}; rc 137 = killed by the timeout.
+set -u
+f="$1"; out="$2"; mkdir -p "$out"
+HERE="$(cd "$(dirname "$0")" && pwd)"
+: "${PROBE:?PROBE must name the conformance-probe binary}"
+: "${PY:?PY must name a python with h5py}"
+TMO="${TMO:-20}"
+MEM_KB="${MEM_KB:-4194304}"
+run() { # name cmd...
+  local name=$1; shift
+  ( ulimit -v "$MEM_KB"; ulimit -c 0; RUST_BACKTRACE=1 exec timeout -s KILL "$TMO" "$@" ) \
+      >"$out/$name.json" 2>"$out/$name.err"
+  echo $? >"$out/$name.rc"
+}
+run ours "$PROBE" "$f"
+run ref "$PY" "$HERE/ref.py" "$f"
+if [ -n "${WITH_H5DUMP:-}" ]; then
+  run h5dump h5dump "$f"
+  : >"$out/h5dump.json" # h5dump's text dump is not compared, only its exit status
+fi
+exit 0

conformance/test_ref.py

@@ -0,0 +1,92 @@
+#!/usr/bin/env python3
+"""Tests of the reference side's corrections: `python conformance/test_ref.py`.
+
+- ref.py compares a big-endian VL sequence by the file's values even though
+  h5py returns them byte-swapped (and records that it corrected them);
+- ref_bugs.py confirms an object only when its reads disagree.
+"""
+import json
+import os
+import subprocess
+import sys
+import tempfile
+import unittest
+
+import h5py
+import numpy as np
+
+HERE = os.path.dirname(os.path.abspath(__file__))
+sys.path.insert(0, HERE)
+import ref_bugs  # noqa: E402
+
+
+def ref_objects(path):
+    out = subprocess.run([sys.executable, os.path.join(HERE, "ref.py"), path],
+                         capture_output=True, text=True, check=True).stdout
+    return {o["path"]: o for o in json.loads(out)["objects"]}
+
+
+class BigEndianVlen(unittest.TestCase):
+    def test_be_vlen_compared_by_file_values(self):
+        with tempfile.TemporaryDirectory() as d:
+            path = os.path.join(d, "v.h5")
+            with h5py.File(path, "w") as f:
+                for name, order in (("be", ">"), ("le", "<")):
+                    t = np.dtype(order + "f4")
+                    ds = f.create_dataset(name, (2,), dtype=h5py.vlen_dtype(t))
+                    ds[0] = np.array([1.0, 2.0], dtype=t)
+                    ds[1] = np.array([3.0], dtype=t)
+                    u = np.dtype(order + "u8")
+                    f.attrs.create(name, [np.array([1, 2], dtype=u), np.array([42], dtype=u)],
+                                   dtype=h5py.vlen_dtype(u))
+            objs = ref_objects(path)
+        be, le = objs["/be"], objs["/le"]
+        # Same values, so the same canonical hash whatever the file's byte order.
+        self.assertEqual(be["hash"], le["hash"])
+        self.assertEqual(objs["/"]["attrs"]["be"]["hash"], objs["/"]["attrs"]["le"]["hash"])
+        self.assertNotIn("ref_fix", le)
+        # And the correction is recorded wherever h5py needed it.
+        import ref
+        if ref.be_vlen_bug():
+            self.assertEqual(be.get("ref_fix"), ["h5py-be-vlen"])
+            self.assertEqual(objs["/"]["attrs"]["be"].get("ref_fix"), ["h5py-be-vlen"])
+
+
+class RefBugsConfirmation(unittest.TestCase):
+    def run_check(self, outcomes):
+        seq = iter(outcomes)
+        saved = ref_bugs.read_once
+        ref_bugs.read_once = lambda *a: next(seq)
+        try:
+            key = next(iter(ref_bugs.READ_BUGS))
+            with tempfile.TemporaryDirectory() as d:
+                p = os.path.join(d, key[0])
+                os.makedirs(os.path.dirname(p))
+                open(p, "wb").close()
+                return ref_bugs.check(d, key)
+        finally:
+            ref_bugs.read_once = saved
+
+    def test_stable_values_are_not_confirmed(self):
+        r = self.run_check(["values a"] * len(ref_bugs.RUNS))
+        self.assertFalse(r["confirmed"])
+
+    def test_changing_values_are_confirmed(self):
+        r = self.run_check(["values a"] * (len(ref_bugs.RUNS) - 1) + ["values b"])
+        self.assertTrue(r["confirmed"])
+        r = self.run_check(["values a"] * (len(ref_bugs.RUNS) - 1) + ["error filter failed"])
+        self.assertTrue(r["confirmed"])
+
+    def test_errors_only_are_not_confirmed(self):
+        # h5py cannot read it at all: nothing it reads, nothing to excuse.
+        r = self.run_check(["error x"] * (len(ref_bugs.RUNS) - 1) + ["error y"])
+        self.assertFalse(r["confirmed"])
+
+    def test_missing_file_is_not_confirmed(self):
+        key = next(iter(ref_bugs.READ_BUGS))
+        with tempfile.TemporaryDirectory() as d:
+            self.assertFalse(ref_bugs.check(d, key)["confirmed"])
+
+
+if __name__ == "__main__":
+    unittest.main()

crates/clawhdf5-accel/Cargo.toml

@@ -1,8 +1,9 @@
 [package]
 name = "clawhdf5-accel"
-version = "2.5.0"
+version = "2.7.0"
 edition = "2024"
-description = "SIMD-accelerated operations for rustyhdf5"
+rust-version.workspace = true
+description = "SIMD kernels (AVX2, NEON) used by clawhdf5 — pure Rust"
 license = "MIT"
 repository = "https://git.redclaw.dev/quantumclaw/clawhdf5"
 readme = "README.md"

crates/clawhdf5-accel/README.md

@@ -1,24 +1,62 @@
 # clawhdf5-accel
 
-[![crates.io](https://img.shields.io/crates/v/clawhdf5-accel.svg)](https://crates.io/crates/clawhdf5-accel)
-[![docs.rs](https://docs.rs/clawhdf5-accel/badge.svg)](https://docs.rs/clawhdf5-accel)
+CPU SIMD kernels for vector search: dot products, cosine similarity, L2
+distance, norms and int8 dot products, dispatched at run time to the best
+backend the CPU has, with a portable scalar fallback for every operation.
+[`clawhdf5-ann`](../clawhdf5-ann/README.md) and
+[`clawhdf5-agent`](../clawhdf5-agent/README.md) use it in their distance
+loops; it has nothing to do with HDF5 file I/O.
 
-SIMD-accelerated operations for clawhdf5.
+Not on crates.io yet; depend on it from git:
+
+```toml
+[dependencies]
+clawhdf5-accel = { git = "https://git.redclaw.dev/quantumclaw/clawhdf5" }
+```
+
+## API
+
+```rust
+use clawhdf5_accel::{cosine_similarity, detect_backend, dot_i8, dot_product, l2_distance};
+
+let a = [1.0f32, 2.0, 3.0, 4.0];
+let b = [4.0f32, 3.0, 2.0, 1.0];
+assert_eq!(dot_product(&a, &b), 20.0);
+let _cos = cosine_similarity(&a, &b);
+let _l2 = l2_distance(&a, &b);
+assert_eq!(dot_i8(&[1, -2, 3], &[4, 5, -6]), -24);
+println!("{:?}", detect_backend()); // e.g. Avx2 on x86-64, Neon on aarch64
+```
+
+Also `vector_norm`, `batch_norms`, `batch_cosine`, `batch_cosine_prenorm`,
+`f16_to_f32_batch`, `checksum_fletcher32` and `align_to_cache_line`.
+
+## Backends
+
+`detect_backend()` picks once per process: `Avx512` (with the `avx512`
+feature), `Avx2` (AVX2 + FMA), `Neon` (every aarch64 CPU), or `Scalar`.
+`Sse4` and `WasmSimd128` are reported when detected but run the scalar
+kernels.
+`dot_i8`, used by the agent's quantised (int8) HNSW index, runs on
+AVX2 and on NEON — with the `SDOT` instruction (through inline assembly,
+since the intrinsic is unstable) on cores that have dotprod, such as the
+Raspberry Pi 5, and plain NEON on older ones. At equal recall the int8
+index answers 1.63x the queries per second of the f32 one on x86-64
+(AVX2; 2026-09-20, machine not recorded, not re-run) and 1.18x on a
+Raspberry Pi 5 (2026-09-21) ([`BENCHMARKS.md` § Quantising the index copy](../../BENCHMARKS.md#quantising-the-index-copy-quantized_index)).
+
+The aarch64 code is compiled out on x86, so only the `test-arm64` CI job
+builds and tests it.
 
 ## Features
 
-- AVX2 and NEON SIMD acceleration
-- AVX-512 support (`avx512` feature)
-- Float16 conversion (`float16` feature)
-- CRC32 checksum acceleration
+| Feature | Default | What | Builds C |
+|---|---|---|---|
+| `avx512` | no | AVX-512F kernels | no |
+| `float16` | no | `f16_to_f32_batch` through the `half` crate (a software conversion otherwise) | no |
 
-## Usage
-
-```rust
-use clawhdf5_accel::checksum::crc32_simd;
-
-let crc = crc32_simd(&data);
-```
+The half-precision conversion used for stored embeddings is
+`clawhdf5_format::float16`, not this crate's.
 
 ## License
 

crates/clawhdf5-accel/src/avx2.rs

@@ -25,6 +25,55 @@ unsafe fn hsum_256(v: __m256) -> f32 {
     _mm_cvtss_f32(result)
 }
 
+/// AVX2 dot product of two `i8` slices, widened to `i32`.
+///
+/// Each 16-byte half is sign-extended to sixteen `i16` lanes and multiplied
+/// pairwise with `madd_epi16`, which sums adjacent products straight into
+/// eight `i32` lanes — the widening that an autovectorised scalar loop does
+/// in several shuffles is one instruction here. A pair sum is at most
+/// `2 * 127 * 127`, far inside `i32`.
+///
+/// # Safety
+/// Caller must verify is_x86_feature_detected!("avx2").
+// SAFETY: Caller must have verified AVX2 via is_x86_feature_detected!.
+#[target_feature(enable = "avx2")]
+pub unsafe fn dot_i8(a: &[i8], b: &[i8]) -> i32 {
+    // SAFETY: Caller guarantees AVX2 is available per the # Safety contract;
+    // every load reads 32 bytes at an index checked against `len` first.
+    unsafe {
+        assert_eq!(a.len(), b.len());
+        let len = a.len();
+        let mut i = 0;
+        let mut acc0 = _mm256_setzero_si256();
+        let mut acc1 = _mm256_setzero_si256();
+
+        while i + 32 <= len {
+            let va = _mm256_loadu_si256(a.as_ptr().add(i).cast());
+            let vb = _mm256_loadu_si256(b.as_ptr().add(i).cast());
+            let a_lo = _mm256_cvtepi8_epi16(_mm256_castsi256_si128(va));
+            let b_lo = _mm256_cvtepi8_epi16(_mm256_castsi256_si128(vb));
+            let a_hi = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(va, 1));
+            let b_hi = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(vb, 1));
+            acc0 = _mm256_add_epi32(acc0, _mm256_madd_epi16(a_lo, b_lo));
+            acc1 = _mm256_add_epi32(acc1, _mm256_madd_epi16(a_hi, b_hi));
+            i += 32;
+        }
+
+        // Horizontal sum of the eight i32 lanes.
+        let v = _mm256_add_epi32(acc0, acc1);
+        let s128 = _mm_add_epi32(_mm256_castsi256_si128(v), _mm256_extracti128_si256(v, 1));
+        let s64 = _mm_add_epi32(s128, _mm_unpackhi_epi64(s128, s128));
+        let s32 = _mm_add_epi32(s64, _mm_shuffle_epi32(s64, 0b01));
+        let mut sum = _mm_cvtsi128_si32(s32);
+
+        while i < len {
+            sum += i32::from(a[i]) * i32::from(b[i]);
+            i += 1;
+        }
+        sum
+    }
+}
+
 /// AVX2 dot product for f32 slices.
 ///
 /// # Safety

crates/clawhdf5-accel/src/lib.rs

@@ -122,6 +122,36 @@ pub fn dot_product(a: &[f32], b: &[f32]) -> f32 {
     }
 }
 
+/// Dot product of two `i8` slices, widened to `i32`.
+///
+/// The kernel behind int8-quantised vector search. On x86-64 it uses the AVX2
+/// path whenever AVX2 is present (including on AVX-512 machines, where it is
+/// what the f32 kernels use too on a default build). On aarch64 it uses the
+/// ARMv8.2 `SDOT` instruction when the CPU has the dot-product extension, and
+/// plain NEON otherwise.
+pub fn dot_i8(a: &[i8], b: &[i8]) -> i32 {
+    match detect_backend() {
+        #[cfg(target_arch = "aarch64")]
+        Backend::Neon => {
+            if std::arch::is_aarch64_feature_detected!("dotprod") {
+                // SAFETY: the dotprod extension was just detected at runtime.
+                unsafe { neon::dot_i8_dotprod(a, b) }
+            } else {
+                // SAFETY: NEON is always available on aarch64.
+                unsafe { neon::dot_i8(a, b) }
+            }
+        }
+
+        #[cfg(target_arch = "x86_64")]
+        // SAFETY: both variants imply AVX2 was detected at runtime (the
+        // AVX-512 backend is only selected on CPUs that also have AVX2).
+        Backend::Avx2 | Backend::Avx512 if is_x86_feature_detected!("avx2") => unsafe {
+            avx2::dot_i8(a, b)
+        },
+        _ => scalar::dot_i8(a, b),
+    }
+}
+
 /// Compute the L2 norm (magnitude) of a vector.
 pub fn vector_norm(v: &[f32]) -> f32 {
     dot_product(v, v).sqrt()
@@ -207,7 +237,10 @@ pub fn f16_to_f32_batch(input: &[u16], output: &mut [f32]) {
     convert::f16_to_f32_batch(input, output);
 }
 
-/// Compute Fletcher-32 checksum.
+/// Compute a textbook Fletcher-32 checksum (both sums start at 0xffff).
+///
+/// This is not HDF5's checksum; the Fletcher-32 I/O filter uses
+/// `clawhdf5_format::checksum::fletcher32`.
 pub fn checksum_fletcher32(data: &[u8]) -> u32 {
     checksum::checksum_fletcher32(data)
 }
@@ -713,3 +746,78 @@ mod tests {
         }
     }
 }
+
+#[cfg(test)]
+mod dot_i8_tests {
+    use super::*;
+
+    fn codes(n: usize, seed: u64) -> Vec<i8> {
+        let mut state = seed;
+        (0..n)
+            .map(|_| {
+                state = state
+                    .wrapping_mul(6_364_136_223_846_793_005)
+                    .wrapping_add(1_442_695_040_888_963_407);
+                // Full range, including the extremes.
+                ((state >> 56) as u8) as i8
+            })
+            .collect()
+    }
+
+    #[test]
+    fn dispatched_kernel_matches_scalar_exactly() {
+        // Integer arithmetic: the SIMD path must agree bit for bit, at every
+        // length — including ones that are not multiples of the 32-byte block,
+        // which exercise the tail.
+        for len in [0, 1, 7, 31, 32, 33, 63, 64, 100, 384, 385, 1536] {
+            let a = codes(len, 1 + len as u64);
+            let b = codes(len, 1000 + len as u64);
+            assert_eq!(dot_i8(&a, &b), scalar::dot_i8(&a, &b), "len {len}");
+        }
+    }
+
+    /// Dispatch only ever takes one path on a given CPU, so on a machine with
+    /// the dot-product extension the plain-NEON kernel would otherwise go
+    /// untested. Check each aarch64 kernel against scalar directly.
+    #[cfg(target_arch = "aarch64")]
+    #[test]
+    fn every_aarch64_kernel_matches_scalar_exactly() {
+        for len in [0, 1, 7, 15, 16, 17, 31, 32, 33, 63, 64, 100, 384, 385, 1536] {
+            let a = codes(len, 7 + len as u64);
+            let b = codes(len, 7000 + len as u64);
+            let want = scalar::dot_i8(&a, &b);
+            // SAFETY: NEON is always available on aarch64.
+            assert_eq!(unsafe { neon::dot_i8(&a, &b) }, want, "neon, len {len}");
+            if std::arch::is_aarch64_feature_detected!("dotprod") {
+                // SAFETY: the dotprod extension was just detected.
+                assert_eq!(
+                    unsafe { neon::dot_i8_dotprod(&a, &b) },
+                    want,
+                    "dotprod, len {len}"
+                );
+            }
+        }
+        // The extremes, through both kernels.
+        let lo = vec![-128i8; 4096];
+        let hi = vec![127i8; 4096];
+        // SAFETY: NEON is always available on aarch64.
+        assert_eq!(unsafe { neon::dot_i8(&lo, &lo) }, 4096 * 128 * 128);
+        // SAFETY: NEON is always available on aarch64.
+        assert_eq!(unsafe { neon::dot_i8(&lo, &hi) }, -4096 * 128 * 127);
+        if std::arch::is_aarch64_feature_detected!("dotprod") {
+            // SAFETY: the dotprod extension was just detected.
+            assert_eq!(unsafe { neon::dot_i8_dotprod(&lo, &lo) }, 4096 * 128 * 128);
+            // SAFETY: the dotprod extension was just detected.
+            assert_eq!(unsafe { neon::dot_i8_dotprod(&lo, &hi) }, -4096 * 128 * 127);
+        }
+    }
+
+    #[test]
+    fn extremes_do_not_overflow() {
+        // -128 * -128 is the largest product; a long run of it must still fit.
+        let a = vec![-128i8; 4096];
+        assert_eq!(dot_i8(&a, &a), 4096 * 128 * 128);
+        let b = vec![127i8; 4096];
+        assert_eq!(dot_i8(&a, &b), -4096 * 128 * 127);
+    }
+}

crates/clawhdf5-accel/src/neon.rs

@@ -180,3 +180,130 @@ pub fn checksum_fletcher32(data: &[u8]) -> u32 {
 
     (sum2 << 16) | sum1
 }
+
+/// NEON dot product of two `i8` slices, widened to `i32`, for any aarch64 CPU.
+///
+/// `vmull_s8` multiplies eight lanes into `i16` — even `-128 * -128` is 16 384,
+/// inside `i16` — and `vpadalq_s16` adds adjacent pairs of those into `i32`
+/// accumulators, so nothing can overflow before the final horizontal sum.
+///
+/// CPUs with the ARMv8.2 dot-product extension should use
+/// [`dot_i8_dotprod`], which does the multiply and the accumulate in one
+/// instruction.
+///
+/// # Safety
+/// Caller must ensure aarch64 target (NEON always available).
+// SAFETY: NEON is always available on aarch64 targets; caller guarantees aarch64.
+#[target_feature(enable = "neon")]
+pub unsafe fn dot_i8(a: &[i8], b: &[i8]) -> i32 {
+    assert_eq!(a.len(), b.len());
+    let len = a.len();
+    let mut i = 0;
+    let mut acc0 = vdupq_n_s32(0);
+    let mut acc1 = vdupq_n_s32(0);
+
+    while i + 16 <= len {
+        // SAFETY: NEON is available per the # Safety contract, and both
+        // 16-byte loads start at an index checked against `len` above.
+        unsafe {
+            let va = vld1q_s8(a.as_ptr().add(i));
+            let vb = vld1q_s8(b.as_ptr().add(i));
+            acc0 = vpadalq_s16(acc0, vmull_s8(vget_low_s8(va), vget_low_s8(vb)));
+            acc1 = vpadalq_s16(acc1, vmull_high_s8(va, vb));
+        }
+        i += 16;
+    }
+
+    let mut sum = vaddvq_s32(vaddq_s32(acc0, acc1));
+    while i < len {
+        sum += i32::from(a[i]) * i32::from(b[i]);
+        i += 1;
+    }
+    sum
+}
+
+/// One `SDOT`: for each of the four `i32` lanes of `acc`, add the dot
+/// product of the corresponding four `i8` pairs from `a` and `b`.
+///
+/// Written as inline assembly because the `vdotq_s32` intrinsic is still
+/// behind the unstable `stdarch_neon_dotprod` feature; inline assembly is
+/// stable on aarch64.
+///
+/// # Safety
+/// Caller must ensure the CPU supports the `dotprod` extension.
+#[inline]
+#[target_feature(enable = "neon,dotprod")]
+unsafe fn sdot(acc: int32x4_t, a: int8x16_t, b: int8x16_t) -> int32x4_t {
+    let mut acc = acc;
+    // SAFETY: `dotprod` is enabled for this function and the caller
+    // guarantees the CPU supports it. The instruction reads only its three
+    // vector registers and touches no memory.
+    unsafe {
+        std::arch::asm!(
+            "sdot {acc:v}.4s, {a:v}.16b, {b:v}.16b",
+            acc = inout(vreg) acc,
+            a = in(vreg) a,
+            b = in(vreg) b,
+            options(pure, nomem, nostack),
+        );
+    }
+    acc
+}
+
+/// NEON dot product of two `i8` slices using the ARMv8.2 dot-product
+/// extension (`SDOT`): sixteen multiply-accumulates per instruction, straight
+/// into `i32` lanes.
+///
+/// Present on the cores this crate actually runs on — Cortex-A76 and later
+/// (Raspberry Pi 5, current Android phones), Neoverse-N1 (Graviton2, Ampere
+/// Altra), and every Apple Silicon generation.
+///
+/// # Safety
+/// Caller must verify `is_aarch64_feature_detected!("dotprod")`.
+// SAFETY: caller has verified the dotprod extension at runtime.
+#[target_feature(enable = "neon,dotprod")]
+pub unsafe fn dot_i8_dotprod(a: &[i8], b: &[i8]) -> i32 {
+    assert_eq!(a.len(), b.len());
+    let len = a.len();
+    let mut i = 0;
+    let mut acc0 = vdupq_n_s32(0);
+    let mut acc1 = vdupq_n_s32(0);
+
+    // Two independent accumulators so consecutive SDOTs are not serialised on
+    // one register.
+    while i + 32 <= len {
+        // SAFETY: dotprod is available per the # Safety contract, and every
+        // 16-byte load starts at an index checked against `len` above.
+        unsafe {
+            acc0 = sdot(
+                acc0,
+                vld1q_s8(a.as_ptr().add(i)),
+                vld1q_s8(b.as_ptr().add(i)),
+            );
+            acc1 = sdot(
+                acc1,
+                vld1q_s8(a.as_ptr().add(i + 16)),
+                vld1q_s8(b.as_ptr().add(i + 16)),
+            );
+        }
+        i += 32;
+    }
+    if i + 16 <= len {
+        // SAFETY: as above; the load is bounds-checked by this condition.
+        unsafe {
+            acc0 = sdot(
+                acc0,
+                vld1q_s8(a.as_ptr().add(i)),
+                vld1q_s8(b.as_ptr().add(i)),
+            );
+        }
+        i += 16;
+    }
+
+    let mut sum = vaddvq_s32(vaddq_s32(acc0, acc1));
+    while i < len {
+        sum += i32::from(a[i]) * i32::from(b[i]);
+        i += 1;
+    }
+    sum
+}

crates/clawhdf5-accel/src/scalar.rs

@@ -140,3 +140,33 @@ fn f16_to_f32_soft(h: u16) -> f32 {
 
     f32::from_bits(f32_bits)
 }
+
+/// Dot product of two `i8` slices, widened to `i32`.
+///
+/// `dim` terms of at most `127 * 127` fit an `i32` for any realistic
+/// dimension (over 130 000 terms before overflow is possible).
+pub fn dot_i8(a: &[i8], b: &[i8]) -> i32 {
+    assert_eq!(a.len(), b.len());
+    // Four independent accumulators over 32-lane blocks: the widening product
+    // has to sit in a fixed-length chunk for the vectoriser to see it, and the
+    // separate accumulators keep it off one dependency chain.
+    const LANE: usize = 8;
+    let (a_blocks, a_tail) = a.as_chunks::<{ LANE * 4 }>();
+    let (b_blocks, b_tail) = b.as_chunks::<{ LANE * 4 }>();
+    let mut acc = [0i32; 4];
+    for (x, y) in a_blocks.iter().zip(b_blocks) {
+        for (lane, slot) in acc.iter_mut().enumerate() {
+            let mut sum = 0i32;
+            for k in 0..LANE {
+                sum += i32::from(x[lane * LANE + k]) * i32::from(y[lane * LANE + k]);
+            }
+            *slot += sum;
+        }
+    }
+    let tail: i32 = a_tail
+        .iter()
+        .zip(b_tail)
+        .map(|(&x, &y)| i32::from(x) * i32::from(y))
+        .sum();
+    acc[0] + acc[1] + acc[2] + acc[3] + tail
+}

crates/clawhdf5-agent/Cargo.toml

@@ -1,7 +1,8 @@
 [package]
 name = "clawhdf5-agent"
-version = "2.5.0"
+version = "2.7.0"
 edition = "2024"
+rust-version.workspace = true
 description = "HDF5-backed persistent memory store for on-device AI agents"
 license = "MIT"
 repository = "https://git.redclaw.dev/quantumclaw/clawhdf5"
@@ -10,14 +11,18 @@ keywords = ["agent", "memory", "hdf5", "vector-search", "embedding"]
 categories = ["database", "science", "algorithms"]
 
 [dependencies]
-clawhdf5-format = { path = "../clawhdf5-format", version = "2.5.0", features = ["parallel", "fast-checksum"] }
-clawhdf5 = { path = "../clawhdf5", version = "2.5.0" }
-clawhdf5-io = { path = "../clawhdf5-io", version = "2.5.0", features = ["mmap"] }
-clawhdf5-accel = { path = "../clawhdf5-accel", version = "2.5.0" }
-clawhdf5-ann = { path = "../clawhdf5-ann", version = "2.5.0", optional = true }
-clawhdf5-gpu = { path = "../clawhdf5-gpu", version = "2.5.0", optional = true, default-features = false }
+clawhdf5-format = { path = "../clawhdf5-format", version = "2.7.0", features = ["parallel", "fast-checksum"] }
+clawhdf5 = { path = "../clawhdf5", version = "2.7.0" }
+clawhdf5-io = { path = "../clawhdf5-io", version = "2.7.0", features = ["mmap"] }
+clawhdf5-accel = { path = "../clawhdf5-accel", version = "2.7.0" }
+clawhdf5-ann = { path = "../clawhdf5-ann", version = "2.7.0", optional = true }
+clawhdf5-gpu = { path = "../clawhdf5-gpu", version = "2.7.0", optional = true, default-features = false }
 serde = { workspace = true }
 byteorder = "1"
+# Signed checkpoints (MemoryConfig-independent; see `signing`). Pure Rust.
+ed25519-dalek = { version = "2", features = ["rand_core"] }
+sha2 = "0.10"
+rand_core = { version = "0.6", features = ["getrandom"] }
 half = { workspace = true, optional = true }
 rayon = { version = "1", optional = true }
 matrixmultiply = { version = "0.3", optional = true }
@@ -44,6 +49,10 @@ harness = false
 name = "memory_bench"
 harness = false
 
+[[bench]]
+name = "multimodal_bench"
+harness = false
+
 [features]
 default = ["float16", "hnsw", "parallel"]
 float16 = ["half"]
@@ -59,7 +68,6 @@ zstd = ["clawhdf5/zstd"]
 # `--no-default-features` (plus re-enabling other defaults) to force the exact
 # linear cosine scan.
 hnsw = ["clawhdf5-ann"]
-agent = []
 gpu = ["clawhdf5-gpu/gpu-wgpu"]
 fast-math = ["matrixmultiply"]
 accelerate = ["accelerate-src", "cblas-sys"]

crates/clawhdf5-agent/README.md

@@ -1,28 +1,124 @@
 # clawhdf5-agent
 
-[![crates.io](https://img.shields.io/crates/v/clawhdf5-agent.svg)](https://crates.io/crates/clawhdf5-agent)
-[![docs.rs](https://img.shields.io/docsrs/clawhdf5-agent)](https://docs.rs/clawhdf5-agent)
+Persistent memory for AI agents in a single HDF5 file: text chunks with
+embeddings and metadata, hybrid search (HNSW vector search + BM25 keyword
+search, fused), sessions, a knowledge graph, a write-ahead log for crash
+safety, and optionally Ed25519-signed checkpoints. Stores open in h5py like
+any other HDF5 file. Built on [`clawhdf5`](../clawhdf5/README.md),
+[`clawhdf5-ann`](../clawhdf5-ann/README.md) and
+[`clawhdf5-accel`](../clawhdf5-accel/README.md).
 
-HDF5-backed persistent memory store for on-device AI agents.
+It is a library: no agent framework integrates it (OpenClaw and ZeroClaw
+integration claims were withdrawn on 2026-09-25; see
+[`docs/openclaw.md`](../../docs/openclaw.md)). The command-line front end
+is [`clawhdf5-cli`](../clawhdf5-cli/README.md).
 
-Built on [clawhdf5](https://crates.io/crates/clawhdf5), clawhdf5-agent provides a vector-searchable memory backend optimized for edge AI workloads. Store embeddings, text chunks, and metadata in a single HDF5 file with SIMD-accelerated similarity search.
-
-## Features
-
-- Persistent vector store in HDF5 format
-- Cosine similarity and L2 distance search
-- SIMD-accelerated via clawhdf5-accel (AVX2, NEON)
-- Optional GPU acceleration via clawhdf5-gpu
-- Memory-mapped access for large stores
-- f16 storage support for compact embeddings
-
-## Usage
+Not on crates.io yet; depend on it from git:
 
 ```toml
 [dependencies]
-clawhdf5-agent = "2.1.0"
+clawhdf5-agent = { git = "https://git.redclaw.dev/quantumclaw/clawhdf5" }
 ```
 
+## Usage
+
+```rust,no_run
+use std::path::PathBuf;
+use clawhdf5_agent::{AgentMemory, HDF5Memory, MemoryConfig, MemoryEntry, SearchOptions};
+
+let config = MemoryConfig::new(PathBuf::from("agent.h5"), "my-agent", 384);
+let mut mem = HDF5Memory::create(config)?;
+
+mem.save(MemoryEntry {
+    chunk: "The deploy key rotates every Monday.".into(),
+    embedding: vec![0.01; 384], // from your embedding model
+    source_channel: "chat".into(),
+    timestamp: 1_790_000_000.0,
+    session_id: "s1".into(),
+    tags: "ops".into(),
+})?;
+
+let query = vec![0.01f32; 384];
+let hits = mem.search(&query, "deploy key", &SearchOptions::new(5).with_sources(["chat"]));
+for h in &hits {
+    println!("{:.3} {}", h.score, h.chunk);
+}
+mem.flush_wal()?; // checkpoint now; otherwise one is made once the WAL holds more than 500 entries (wal_max_entries)
+# Ok::<(), clawhdf5_agent::MemoryError>(())
+```
+
+## What is in it
+
+- **`HDF5Memory`** — `create`, `open` (single writer: an exclusive lock on
+  `<store>.h5.lock`, a second opener gets `MemoryError::Locked`),
+  `open_read_only` (no lock, never writes). Through the `AgentMemory`
+  trait: `save`, `save_batch`, `delete`, `compact`, `count`, `snapshot`,
+  sessions; also `save_or_update`, `delete_batch`, `flush_wal`.
+- **Search** — `search(query_embedding, text, &SearchOptions)`: optional
+  source-channel filter applied before ranking, vector + BM25 fusion
+  (weighted or RRF), Hebbian activation scaling, optional re-ranking
+  (`reranker::ReRankConfig`) and confidence rejection
+  (`confidence::ConfidenceConfig`). `hybrid_search` and
+  `hybrid_search_with` are thin wrappers. The vector stage uses the HNSW
+  index (`hnsw` feature); its graph is saved to `<store>.h5.ann` at each
+  checkpoint and reloaded on open (rebuilt if stale or damaged).
+- **Storage settings** (`MemoryConfig`, persisted with the store):
+  `float16` embeddings (on by default for new stores; 48% smaller file at
+  100K records, same retrieval on LongMemEval), `quantized_index` (int8
+  copy of the vectors in the index, on by default; re-scored against the
+  exact embeddings), `compression` (off by default), HNSW `m`/`ef`
+  parameters, WAL settings (`wal_enabled`, on by default; `wal_max_entries`,
+  500: the WAL is checkpointed into the `.h5` once it holds more).
+- **WAL** (`wal`) — every write is appended to `<store>.h5.wal` with a
+  chained CRC32 per entry, so a corrupted, reordered or spliced entry stops
+  replay. Recovers from a process crash at any point, including between a
+  checkpoint and the WAL truncate. WAL appends are not fsynced: saves since
+  the last checkpoint can be lost on power failure. An unreadable WAL is
+  quarantined to `<store>.h5.wal.corrupt-<ts>`.
+- **Signed checkpoints** (`signing`) — `set_signing_key` signs a manifest
+  (SHA-256 Merkle tree over records, plus settings, sessions and graph) at
+  every checkpoint; `HDF5Memory::verify(path, &public_key)` checks it and
+  locates edits. WAL entries after the checkpoint are not covered.
+- **Knowledge graph** (`knowledge`, `entity_extract`) — `add_entity`,
+  `add_entity_alias`, `add_relation`, `extract_and_store_entities`,
+  traversal and spreading activation.
+- **Also:** sessions (`session`), temporal index (`temporal`),
+  consolidation tiers (`consolidation`), an in-memory TTL tier
+  (`ephemeral`), multi-modal embeddings (`multimodal`), `AGENTS.md`
+  generation (`agents_md`), query expansion, and a session-scoped
+  provenance ledger and write-anomaly detector on every save
+  (`take_anomaly_alerts`; alerts never block a save, and the source is
+  inferred from `source_channel`, not authenticated).
+- `openclaw::ClawhdfBackend` is `search` with re-ranking and confidence
+  on, plus Markdown import/export. The module name is historical: it is not
+  an OpenClaw plugin.
+
+## Features
+
+| Feature | Default | What | Builds C |
+|---|---|---|---|
+| `hnsw` | yes | HNSW vector index (`clawhdf5-ann`); without it the vector stage is an exact linear cosine scan | no |
+| `parallel` | yes | build the HNSW index on a rayon pool (same graph either way) | no |
+| `float16` | yes | f16 helpers in `vector_search` (`half`). Stores' `MemoryConfig::float16` works without it. | no |
+| `fast-math` | no | `matrixmultiply` batch distances in `strategy` | no |
+| `accelerate` | no | Apple Accelerate BLAS in `strategy` (macOS) | links a system framework |
+| `openblas` | no | OpenBLAS in `strategy` | yes (`openblas-src`) |
+| `gpu` | no | `gpu_search` through [`clawhdf5-gpu`](../clawhdf5-gpu/README.md) (wgpu), used by `strategy`, not by `HDF5Memory::search` | no, but needs GPU drivers |
+| `zstd` | no | Zstd instead of deflate when `MemoryConfig::compression` is on | yes (libzstd) |
+| `async` | no | `async_memory` wrapper on tokio | no |
+
+`--no-default-features --features float16` forces the exact linear scan.
+
+## Measurements and limits
+
+- Search recall and latency, file size, LongMemEval and MemoryArena
+  retrieval numbers: [`BENCHMARKS.md`](../../BENCHMARKS.md), measured with
+  the `clawhdf5-bench` binaries (`search_harness`, `longmemeval_bench`,
+  `footprint_bench`, ...).
+- Known issues and their history: [`docs/known-issues.md`](../../docs/known-issues.md).
+- Migrating a SQLite memory database:
+  [`clawhdf5-migrate`](../clawhdf5-migrate/README.md).
+
 ## License
 
 MIT

crates/clawhdf5-agent/benches/multimodal_bench.rs

@@ -0,0 +1,107 @@
+//! Multi-modal memory search benchmarks (`clawhdf5_agent::multimodal`).
+//!
+//! Covers `MultiModalStore::search_cross_modal` (every embedding of every
+//! record, whatever its modality) and, for comparison,
+//! `MultiModalStore::search_by_modality` restricted to one modality.
+//!
+//! Corpus: N records (1K and 10K), each carrying two 384-dim embeddings —
+//! a text embedding of its caption plus one embedding of its primary modality,
+//! cycling Image / Audio / Video — so a cross-modal query scores 2N vectors.
+//! All data comes from a fixed-seed LCG, so every run sees the same corpus.
+//!
+//! Run: `cargo bench -p clawhdf5-agent --bench multimodal_bench`
+
+use std::collections::HashMap;
+
+use clawhdf5_agent::multimodal::{
+    MediaRef, ModalEmbedding, Modality, MultiModalRecord, MultiModalStore,
+};
+use criterion::{BenchmarkId, Criterion, criterion_group, criterion_main};
+
+// ---------------------------------------------------------------------------
+// Simple deterministic PRNG (LCG), same as the other agent benches
+// ---------------------------------------------------------------------------
+
+struct Rng(u32);
+
+impl Rng {
+    fn new(seed: u32) -> Self {
+        Self(seed)
+    }
+    fn next_u32(&mut self) -> u32 {
+        self.0 = self.0.wrapping_mul(1103515245).wrapping_add(12345);
+        self.0 >> 16
+    }
+    fn next_f32(&mut self) -> f32 {
+        self.next_u32() as f32 / 65536.0 - 0.5
+    }
+}
+
+fn make_vec(rng: &mut Rng, dim: usize) -> Vec<f32> {
+    (0..dim).map(|_| rng.next_f32()).collect()
+}
+
+// ---------------------------------------------------------------------------
+// Corpus
+// ---------------------------------------------------------------------------
+
+const DIM: usize = 384;
+const K: usize = 10;
+
+const MEDIA: [(Modality, &str, &str); 3] = [
+    (Modality::Image, "image/png", "clip-vit-base"),
+    (Modality::Audio, "audio/wav", "clap-base"),
+    (Modality::Video, "video/mp4", "xclip-base"),
+];
+
+fn build_store(n: usize, seed: u32) -> MultiModalStore {
+    let mut rng = Rng::new(seed);
+    let mut store = MultiModalStore::new();
+    for i in 0..n {
+        let (modality, mime, model) = &MEDIA[i % MEDIA.len()];
+        let embeddings = vec![
+            ModalEmbedding::new(Modality::Text, make_vec(&mut rng, DIM), "minilm-l6"),
+            ModalEmbedding::new(modality.clone(), make_vec(&mut rng, DIM), *model),
+        ];
+        store.add_record(MultiModalRecord {
+            id: 0,
+            primary_modality: modality.clone(),
+            text_content: Some(format!("{modality} memory {i}")),
+            media_ref: Some(MediaRef::path(format!("/media/{i}"), *mime)),
+            embeddings,
+            observation: None,
+            timestamp: 1_700_000_000.0 + i as f64,
+            metadata: HashMap::new(),
+        });
+    }
+    store
+}
+
+// ---------------------------------------------------------------------------
+// Benchmarks
+// ---------------------------------------------------------------------------
+
+fn multimodal_search_benches(c: &mut Criterion) {
+    let query = make_vec(&mut Rng::new(99), DIM);
+
+    let mut group = c.benchmark_group("multimodal_search");
+    group.sample_size(50);
+
+    for (label, n) in [("1k", 1_000usize), ("10k", 10_000)] {
+        let store = build_store(n, 42);
+        assert_eq!(store.count(), n);
+
+        group.bench_with_input(BenchmarkId::new("cross_modal", label), &n, |b, _| {
+            b.iter(|| store.search_cross_modal(&query, K));
+        });
+
+        group.bench_with_input(BenchmarkId::new("by_modality_image", label), &n, |b, _| {
+            b.iter(|| store.search_by_modality(&Modality::Image, &query, K));
+        });
+    }
+
+    group.finish();
+}
+
+criterion_group!(multimodal_benches, multimodal_search_benches);
+criterion_main!(multimodal_benches);

crates/clawhdf5-agent/src/agents_md.rs

@@ -118,6 +118,10 @@ mod tests {
             created_at: "2025-01-01T00:00:00Z".to_string(),
             wal_enabled: false,
             wal_max_entries: 500,
+            quantized_index: false,
+            hnsw_m: 16,
+            hnsw_ef_construction: 64,
+            hnsw_ef_search: 0,
         }
     }
 

crates/clawhdf5-agent/src/bm25.rs

@@ -88,8 +88,11 @@ impl BM25Index {
     /// Search the index for a query, returning the top `k` results
     /// as `(doc_id, score)` pairs sorted by score descending.
     ///
-    /// Uses Block-Max WAND for early termination when remaining documents
-    /// cannot beat the current top-k threshold.
+    /// Scores every matching document exhaustively, then keeps the top `k`.
+    /// There is no early termination (WAND, MaxScore): the store's hot path
+    /// is [`scores`](Self::scores), because score fusion normalises over the
+    /// whole matching set and so needs every score, which no pruning scheme
+    /// can skip. This method is for BM25-only callers.
     pub fn search(&self, query: &str, k: usize) -> Vec<(usize, f32)> {
         if k == 0 {
             return Vec::new();
@@ -561,8 +564,9 @@ mod tests {
     }
 
     #[test]
-    fn wand_returns_same_results_as_exhaustive() {
-        // WAND-style search should produce same scores as exhaustive
+    fn top_k_search_matches_ranking_every_score() {
+        // `search` must agree with ranking the full `scores` set — the
+        // bounded heap is an optimisation over sorting, not an approximation.
         let docs: Vec<String> = (0..100)
             .map(|i| {
                 if i % 3 == 0 {

crates/clawhdf5-agent/src/cache.rs

@@ -1,17 +1,145 @@
 //! In-memory cache for memory entries, sessions, and knowledge graph.
 
 use crate::vector_search;
+use clawhdf5_format::float16::round_to_f16;
+
+/// Every entry's embedding, in one contiguous `[N x dim]` buffer.
+///
+/// Rows are always exactly `dim` long: a shorter one is zero-padded, a longer
+/// one truncated. The previous `Vec<Vec<f32>>` allowed ragged rows, which
+/// silently misaligned the flattened copy that the batched kernels read — a
+/// single wrong-length embedding shifted every row after it. Padding makes
+/// that unrepresentable. A record stored without an embedding therefore holds
+/// a zero row, and is told apart by its norm being zero rather than by length.
+///
+/// This used to be two fields — a `Vec<Vec<f32>>` and a flattened copy kept in
+/// lock-step — which stored the whole corpus twice and cost one heap
+/// allocation per entry on top. At 100k 384-dim entries that duplicate was
+/// ~150 MiB. Indexing yields a `&[f32]` row, so `embeddings[i]` still reads
+/// the same way.
+#[derive(Debug, Clone, Default)]
+pub struct Embeddings {
+    flat: Vec<f32>,
+    dim: usize,
+}
+
+impl Embeddings {
+    pub fn new(dim: usize) -> Self {
+        Self {
+            flat: Vec::new(),
+            dim,
+        }
+    }
+
+    /// Number of embeddings.
+    pub fn len(&self) -> usize {
+        self.flat.len().checked_div(self.dim).unwrap_or(0)
+    }
+
+    pub fn is_empty(&self) -> bool {
+        self.len() == 0
+    }
+
+    /// The whole buffer, `[N x dim]` row-major — what batched kernels read.
+    pub fn as_flat(&self) -> &[f32] {
+        &self.flat
+    }
+
+    pub fn dim(&self) -> usize {
+        self.dim
+    }
+
+    /// Row `i`, or `None` if out of range.
+    pub fn get(&self, i: usize) -> Option<&[f32]> {
+        let start = i.checked_mul(self.dim)?;
+        self.flat.get(start..start.checked_add(self.dim)?)
+    }
+
+    pub fn iter(&self) -> impl ExactSizeIterator<Item = &[f32]> {
+        self.flat.chunks_exact(self.dim.max(1))
+    }
+
+    /// Append one embedding. A row whose length doesn't match `dim` is padded
+    /// or truncated, so the buffer stays rectangular whatever a caller passes.
+    pub fn push(&mut self, embedding: &[f32]) {
+        if self.dim == 0 {
+            return;
+        }
+        let take = embedding.len().min(self.dim);
+        self.flat.extend_from_slice(&embedding[..take]);
+        self.flat.resize(self.flat.len() + (self.dim - take), 0.0);
+    }
+
+    /// Replace row `i`. Out-of-range indices are ignored.
+    pub fn set(&mut self, i: usize, embedding: &[f32]) {
+        let Some(start) = i.checked_mul(self.dim) else {
+            return;
+        };
+        if start + self.dim > self.flat.len() {
+            return;
+        }
+        let take = embedding.len().min(self.dim);
+        self.flat[start..start + take].copy_from_slice(&embedding[..take]);
+        self.flat[start + take..start + self.dim].fill(0.0);
+    }
+
+    /// Keep only the rows `keep` returns true for, preserving order.
+    pub fn retain(&mut self, mut keep: impl FnMut(usize) -> bool) {
+        if self.dim == 0 {
+            return;
+        }
+        let mut write = 0usize;
+        for read in 0..self.len() {
+            if keep(read) {
+                if write != read {
+                    let (dst, src) = (write * self.dim, read * self.dim);
+                    self.flat.copy_within(src..src + self.dim, dst);
+                }
+                write += 1;
+            }
+        }
+        self.flat.truncate(write * self.dim);
+    }
+
+    /// Replace the contents with `rows`.
+    pub fn reset_from(&mut self, dim: usize, rows: impl IntoIterator<Item = Vec<f32>>) {
+        self.dim = dim;
+        self.flat.clear();
+        for row in rows {
+            self.push(&row);
+        }
+    }
+
+    /// Adopt an already-flat buffer, trimming any partial trailing row.
+    pub fn set_flat(&mut self, dim: usize, mut flat: Vec<f32>) {
+        self.dim = dim;
+        match flat.len().checked_div(dim) {
+            Some(rows) => flat.truncate(rows * dim),
+            None => flat.clear(),
+        }
+        self.flat = flat;
+    }
+}
+
+impl PartialEq for Embeddings {
+    fn eq(&self, other: &Self) -> bool {
+        self.dim == other.dim && self.flat == other.flat
+    }
+}
+
+impl std::ops::Index<usize> for Embeddings {
+    type Output = [f32];
+
+    fn index(&self, i: usize) -> &[f32] {
+        self.get(i).expect("embedding index out of range")
+    }
+}
 
 /// In-memory cache for the /memory group data.
 #[derive(Debug, Clone)]
 pub struct MemoryCache {
     pub chunks: Vec<String>,
-    pub embeddings: Vec<Vec<f32>>,
-    /// `embeddings` flattened into one contiguous `[N × embedding_dim]`
-    /// buffer, maintained incrementally alongside `embeddings` (push/update/
-    /// compact) so BLAS/Accelerate batch search can read it directly instead
-    /// of re-flattening the whole corpus on every query.
-    pub embeddings_flat: Vec<f32>,
+    pub embeddings: Embeddings,
     pub source_channels: Vec<String>,
     pub timestamps: Vec<f64>,
     pub session_ids: Vec<String>,
@@ -22,14 +150,18 @@ pub struct MemoryCache {
     pub norms: Vec<f32>,
     /// Hebbian activation weights (default 1.0 per entry).
     pub activation_weights: Vec<f32>,
+    /// Round every embedding to IEEE half precision as it enters the cache,
+    /// so the cache holds exactly what a `float16` store writes to disk. Set
+    /// it with [`MemoryCache::set_half_precision`], which also rounds the
+    /// rows already held.
+    pub half_precision: bool,
 }
 
 impl MemoryCache {
     pub fn new(embedding_dim: usize) -> Self {
         Self {
             chunks: Vec::new(),
-            embeddings: Vec::new(),
-            embeddings_flat: Vec::new(),
+            embeddings: Embeddings::new(embedding_dim),
             source_channels: Vec::new(),
             timestamps: Vec::new(),
             session_ids: Vec::new(),
@@ -38,18 +170,52 @@ impl MemoryCache {
             embedding_dim,
             norms: Vec::new(),
             activation_weights: Vec::new(),
+            half_precision: false,
         }
     }
 
-    /// Rebuild `embeddings_flat` from `embeddings` from scratch. Callers that
-    /// populate `embeddings` directly (bulk loads) must call this afterward.
-    pub fn rebuild_flat(&mut self) {
-        self.embeddings_flat.clear();
-        self.embeddings_flat
-            .reserve(self.embeddings.len() * self.embedding_dim);
-        for emb in &self.embeddings {
-            self.embeddings_flat.extend_from_slice(emb);
+    /// Switch half-precision rounding on or off. Turning it on rounds every
+    /// embedding already held (and recomputes norms where one changed) —
+    /// e.g. a `float16` store whose last checkpoint predates half-precision
+    /// storage and so is still `f32` on disk.
+    pub fn set_half_precision(&mut self, on: bool) {
+        self.half_precision = on;
+        if !on {
+            return;
         }
+        for i in 0..self.embeddings.len() {
+            let row = &self.embeddings[i];
+            if row
+                .iter()
+                .all(|&v| round_to_f16(v).to_bits() == v.to_bits())
+            {
+                continue;
+            }
+            let rounded: Vec<f32> = row.iter().map(|&v| round_to_f16(v)).collect();
+            self.norms[i] = vector_search::compute_norm(&rounded);
+            self.embeddings.set(i, &rounded);
+        }
+    }
+
+    /// The embedding as the cache will hold it: rounded to half precision
+    /// when [`Self::half_precision`] is on, otherwise unchanged.
+    fn stored_form(&self, mut embedding: Vec<f32>) -> Vec<f32> {
+        if self.half_precision {
+            for v in &mut embedding {
+                *v = round_to_f16(*v);
+            }
+        }
+        embedding
+    }
+
+    /// Kept for callers that used to have to re-flatten after a bulk load.
+    /// The buffer is always flat now, so there is nothing to rebuild.
+    #[deprecated(note = "embeddings are stored flat; this is a no-op")]
+    pub fn rebuild_flat(&mut self) {}
+
+    /// The embeddings as one contiguous `[N x dim]` buffer.
+    pub fn flat_embeddings(&self) -> &[f32] {
+        self.embeddings.as_flat()
     }
 
     /// Total number of entries (including tombstoned).
@@ -77,10 +243,10 @@ impl MemoryCache {
         tags: String,
     ) -> usize {
         let idx = self.chunks.len();
+        let embedding = self.stored_form(embedding);
         let norm = vector_search::compute_norm(&embedding);
         self.chunks.push(chunk);
-        self.embeddings_flat.extend_from_slice(&embedding);
-        self.embeddings.push(embedding);
+        self.embeddings.push(&embedding);
         self.source_channels.push(source_channel);
         self.timestamps.push(timestamp);
         self.session_ids.push(session_id);
@@ -116,22 +282,10 @@ impl MemoryCache {
         session_id: String,
     ) {
         if idx < self.chunks.len() {
+            let embedding = self.stored_form(embedding);
             let norm = vector_search::compute_norm(&embedding);
             self.chunks[idx] = chunk;
-            let dim = self.embedding_dim;
-            let flat_start = idx * dim;
-            let matches_dim =
-                embedding.len() == dim && flat_start + dim <= self.embeddings_flat.len();
-            self.embeddings[idx] = embedding;
-            if matches_dim {
-                self.embeddings_flat[flat_start..flat_start + dim]
-                    .copy_from_slice(&self.embeddings[idx]);
-            } else {
-                // Embedding length doesn't match embedding_dim (shouldn't
-                // happen in practice) — fall back to a full rebuild rather
-                // than leave embeddings_flat misaligned with embeddings.
-                self.rebuild_flat();
-            }
+            self.embeddings.set(idx, &embedding);
             self.source_channels[idx] = source_channel;
             self.timestamps[idx] = timestamp;
             self.session_ids[idx] = session_id;
@@ -183,7 +337,7 @@ impl MemoryCache {
                 new_idx += 1;
                 let norm = vector_search::compute_norm(&self.embeddings[i]);
                 new_chunks.push(self.chunks[i].clone());
-                new_embeddings.push(self.embeddings[i].clone());
+                new_embeddings.push(self.embeddings[i].to_vec());
                 new_source_channels.push(self.source_channels[i].clone());
                 new_timestamps.push(self.timestamps[i]);
                 new_session_ids.push(self.session_ids[i].clone());
@@ -196,7 +350,8 @@ impl MemoryCache {
 
         let removed = old_len - new_chunks.len();
         self.chunks = new_chunks;
-        self.embeddings = new_embeddings;
+        self.embeddings
+            .reset_from(self.embedding_dim, new_embeddings);
         self.source_channels = new_source_channels;
         self.timestamps = new_timestamps;
         self.session_ids = new_session_ids;
@@ -204,16 +359,14 @@ impl MemoryCache {
         self.tombstones = new_tombstones;
         self.norms = new_norms;
         self.activation_weights = new_activation_weights;
-        self.rebuild_flat();
 
         (removed, index_map)
     }
 
-    /// Flatten all embeddings into a single Vec<f32> for HDF5 storage.
-    /// `embeddings_flat` is already maintained incrementally, so this just
-    /// clones it — kept as a method for callers that want an owned copy.
-    pub fn flat_embeddings(&self) -> Vec<f32> {
-        self.embeddings_flat.clone()
+    /// All embeddings as one owned `[N x dim]` buffer, for HDF5 storage.
+    /// Prefer [`MemoryCache::flat_embeddings`] where a borrow will do.
+    pub fn flat_embeddings_owned(&self) -> Vec<f32> {
+        self.embeddings.as_flat().to_vec()
     }
 }
 
@@ -224,7 +377,7 @@ mod tests {
     /// `embeddings_flat` must always equal a from-scratch flatten of `embeddings`.
     fn assert_flat_in_sync(cache: &MemoryCache) {
         let expected: Vec<f32> = cache.embeddings.iter().flatten().copied().collect();
-        assert_eq!(cache.embeddings_flat, expected);
+        assert_eq!(cache.embeddings.as_flat(), expected);
     }
 
     #[test]
@@ -247,7 +400,10 @@ mod tests {
             String::new(),
         );
         assert_flat_in_sync(&cache);
-        assert_eq!(cache.embeddings_flat, vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0]);
+        assert_eq!(
+            cache.embeddings.as_flat(),
+            vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0]
+        );
     }
 
     #[test]
@@ -279,7 +435,7 @@ mod tests {
         );
         assert_flat_in_sync(&cache);
         assert_eq!(
-            cache.embeddings_flat,
+            cache.embeddings.as_flat(),
             vec![7.0, 8.0, 9.0, 4.0, 5.0, 6.0],
             "update must overwrite the correct flat slice, not just append"
         );
@@ -315,14 +471,71 @@ mod tests {
         cache.mark_deleted(1);
         cache.compact();
         assert_flat_in_sync(&cache);
-        assert_eq!(cache.embeddings_flat, vec![1.0, 1.0, 3.0, 3.0]);
+        assert_eq!(cache.embeddings.as_flat(), vec![1.0, 1.0, 3.0, 3.0]);
     }
 
     #[test]
     fn rebuild_flat_matches_manual_flatten() {
         let mut cache = MemoryCache::new(2);
-        cache.embeddings = vec![vec![1.0, 2.0], vec![3.0, 4.0]];
-        cache.rebuild_flat();
-        assert_eq!(cache.embeddings_flat, vec![1.0, 2.0, 3.0, 4.0]);
+        cache
+            .embeddings
+            .reset_from(2, vec![vec![1.0, 2.0], vec![3.0, 4.0]]);
+        assert_eq!(cache.embeddings.as_flat(), vec![1.0, 2.0, 3.0, 4.0]);
+    }
+
+    #[test]
+    fn set_half_precision_rounds_existing_rows_and_their_norms() {
+        // A store with float16 set whose checkpoint is still f32 on disk
+        // loads full-precision rows; switching rounding on must bring them to
+        // exactly what the next checkpoint will write.
+        let mut cache = MemoryCache::new(3);
+        cache.push(
+            "a".into(),
+            vec![0.1, 0.2, 0.3],
+            "c".into(),
+            0.0,
+            "s".into(),
+            "".into(),
+        );
+        cache.push(
+            "b".into(),
+            vec![0.5, 0.25, 1.0],
+            "c".into(),
+            0.0,
+            "s".into(),
+            "".into(),
+        );
+        let exact_norm = cache.norms[0];
+
+        cache.set_half_precision(true);
+        let row0: Vec<f32> = [0.1f32, 0.2, 0.3]
+            .iter()
+            .map(|&v| round_to_f16(v))
+            .collect();
+        assert_eq!(&cache.embeddings[0], row0.as_slice());
+        assert_eq!(cache.norms[0], vector_search::compute_norm(&row0));
+        assert_ne!(cache.norms[0], exact_norm);
+        // Already representable: untouched.
+        assert_eq!(&cache.embeddings[1], &[0.5, 0.25, 1.0]);
+
+        // New rows are rounded as they arrive, and updates too.
+        cache.push(
+            "c".into(),
+            vec![0.1, 0.0, 0.0],
+            "c".into(),
+            0.0,
+            "s".into(),
+            "".into(),
+        );
+        assert_eq!(cache.embeddings[2][0], round_to_f16(0.1));
+        cache.update(
+            2,
+            "c".into(),
+            vec![0.3, 0.0, 0.0],
+            "c".into(),
+            0.0,
+            "s".into(),
+        );
+        assert_eq!(cache.embeddings[2][0], round_to_f16(0.3));
     }
 }

crates/clawhdf5-agent/src/consolidation.rs

@@ -144,9 +144,44 @@ pub struct ConsolidationStats {
 
 pub struct ImportanceScorer;
 
+/// Sum of squares, in 8-wide lanes so it vectorises.
+fn sum_of_squares(a: &[f32]) -> f32 {
+    let (blocks, tail) = a.as_chunks::<8>();
+    let mut acc = [0.0f32; 8];
+    for b in blocks {
+        for i in 0..8 {
+            acc[i] += b[i] * b[i];
+        }
+    }
+    acc.iter().sum::<f32>() + tail.iter().map(|x| x * x).sum::<f32>()
+}
+
+/// `(a · b, |b|²)` in one pass over equal-length slices, in 8-wide lanes.
+fn dot_and_norm2(a: &[f32], b: &[f32]) -> (f32, f32) {
+    let (a_blocks, a_tail) = a.as_chunks::<8>();
+    let (b_blocks, b_tail) = b.as_chunks::<8>();
+    let mut dot = [0.0f32; 8];
+    let mut nb = [0.0f32; 8];
+    for (x, y) in a_blocks.iter().zip(b_blocks) {
+        for i in 0..8 {
+            dot[i] += x[i] * y[i];
+            nb[i] += y[i] * y[i];
+        }
+    }
+    let mut d = dot.iter().sum::<f32>();
+    let mut n = nb.iter().sum::<f32>();
+    for (x, y) in a_tail.iter().zip(b_tail) {
+        d += x * y;
+        n += y * y;
+    }
+    (d, n)
+}
+
 impl ImportanceScorer {
     /// Cosine similarity between two embedding slices.
-    /// Returns 0.0 if either norm is zero.
+    /// Returns 0.0 if either norm is zero. The reference that
+    /// [`Self::score_surprise`] is tested against.
+    #[cfg(test)]
     fn cosine_similarity(a: &[f32], b: &[f32]) -> f32 {
         let len = a.len().min(b.len());
         if len == 0 {
@@ -167,13 +202,54 @@ impl ImportanceScorer {
 
     /// Novelty score: 1.0 − max cosine similarity against all existing records.
     /// Returns 1.0 when there are no existing memories.
+    ///
+    /// Same result as the reference cosine similarity against each record, but the
+    /// new embedding's norm is computed once rather than per record, each
+    /// record costs one fused pass (dot product and its norm together) rather
+    /// than three, and a large working set is scored in parallel. Every insert
+    /// scores against the whole working tier, so this is what an unbounded
+    /// working tier pays for: at 100K records it was the difference between a
+    /// benchmark finishing and not (`BENCHMARKS.md`, "Consolidation Efficiency").
     pub fn score_surprise(embedding: &[f32], existing_memories: &[&MemoryRecord]) -> f32 {
         if existing_memories.is_empty() {
             return 1.0;
         }
+        let query_norm2 = sum_of_squares(embedding);
+        let similarity = |r: &&MemoryRecord| -> f32 {
+            let other = &r.embedding;
+            let len = embedding.len().min(other.len());
+            if len == 0 {
+                return 0.0;
+            }
+            let (dot, other_norm2) = dot_and_norm2(&embedding[..len], &other[..len]);
+            // A shorter record compares against the query's matching prefix.
+            let q2 = if len == embedding.len() {
+                query_norm2
+            } else {
+                sum_of_squares(&embedding[..len])
+            };
+            if q2 == 0.0 || other_norm2 == 0.0 {
+                return 0.0;
+            }
+            dot / (q2.sqrt() * other_norm2.sqrt())
+        };
+        #[cfg(feature = "parallel")]
+        let max_sim = if existing_memories.len() >= 4096 {
+            use rayon::prelude::*;
+            existing_memories
+                .par_iter()
+                .map(similarity)
+                .reduce(|| f32::NEG_INFINITY, f32::max)
+        } else {
+            existing_memories
+                .iter()
+                .map(similarity)
+                .fold(f32::NEG_INFINITY, f32::max)
+        };
+        #[cfg(not(feature = "parallel"))]
         let max_sim = existing_memories
             .iter()
-            .map(|r| Self::cosine_similarity(embedding, &r.embedding))
+            .map(similarity)
             .fold(f32::NEG_INFINITY, f32::max);
         (1.0 - max_sim).clamp(0.0, 1.0)
     }
@@ -471,6 +547,54 @@ impl ConsolidationEngine {
 mod tests {
     use super::*;
 
+    #[test]
+    fn score_surprise_matches_the_reference_cosine() {
+        let mut x = 0x2545_F491_4F6C_DD1Du64;
+        let mut next = || {
+            x ^= x << 13;
+            x ^= x >> 7;
+            x ^= x << 17;
+            (x >> 40) as f32 / (1u64 << 24) as f32 - 0.5
+        };
+        let make = |id: u64, v: Vec<f32>| MemoryRecord {
+            id,
+            chunk: String::new(),
+            embedding: v,
+            tier: MemoryTier::Working,
+            importance: 0.0,
+            access_count: 0,
+            last_accessed: 0.0,
+            created_at: 0.0,
+            source: MemorySource::User,
+        };
+        // Ordinary rows, a shorter one, an empty one and a zero vector; and
+        // enough rows to take the parallel path too.
+        for n in [5usize, 5000] {
+            let mut recs: Vec<MemoryRecord> = (0..n as u64)
+                .map(|i| make(i, (0..37).map(|_| next()).collect()))
+                .collect();
+            recs.push(make(9_000, (0..20).map(|_| next()).collect()));
+            recs.push(make(9_001, Vec::new()));
+            recs.push(make(9_002, vec![0.0; 37]));
+            let refs: Vec<&MemoryRecord> = recs.iter().collect();
+            for _ in 0..5 {
+                let q: Vec<f32> = (0..37).map(|_| next()).collect();
+                let expected = (1.0
+                    - refs
+                        .iter()
+                        .map(|r| ImportanceScorer::cosine_similarity(&q, &r.embedding))
+                        .fold(f32::NEG_INFINITY, f32::max))
+                .clamp(0.0, 1.0);
+                let got = ImportanceScorer::score_surprise(&q, &refs);
+                assert!((got - expected).abs() < 1e-5, "n={n}: {got} vs {expected}");
+            }
+        }
+        assert_eq!(
+            ImportanceScorer::score_surprise(&[0.0; 4], &[&make(1, vec![1.0; 4])]),
+            1.0
+        );
+    }
+
     // Helper: build a simple normalised embedding of given dimension.
     fn unit_vec(dim: usize, hot: usize) -> Vec<f32> {
         let mut v = vec![0.0f32; dim];

crates/clawhdf5-agent/src/hybrid.rs

@@ -28,7 +28,7 @@ use crate::vector_search;
 pub fn hybrid_search(
     query_embedding: &[f32],
     query_text: &str,
-    vectors: &[Vec<f32>],
+    vectors: &(impl crate::vector_search::VectorSet + Sync + ?Sized),
     chunks: &[String],
     tombstones: &[u8],
     bm25_index: &BM25Index,
@@ -56,7 +56,7 @@ pub fn hybrid_search(
 pub fn hybrid_search_fused(
     query_embedding: &[f32],
     query_text: &str,
-    vectors: &[Vec<f32>],
+    vectors: &(impl crate::vector_search::VectorSet + Sync + ?Sized),
     _chunks: &[String],
     tombstones: &[u8],
     bm25_index: &BM25Index,
@@ -65,31 +65,34 @@ pub fn hybrid_search_fused(
 ) -> Vec<(usize, f32)> {
     // Get raw scores from both systems. Request all results so normalization
     // covers the full distribution.
-    // Use parallel search when rayon feature is enabled and vector count > 10K.
-    let vec_scores = {
-        #[cfg(feature = "parallel")]
-        {
-            if vectors.len() > 10_000 {
-                vector_search::parallel_cosine_batch(
-                    query_embedding,
-                    vectors,
-                    tombstones,
-                    vectors.len(),
-                )
-            } else {
-                vector_search::cosine_similarity_batch(query_embedding, vectors, tombstones)
-            }
-        }
-        #[cfg(not(feature = "parallel"))]
-        {
-            vector_search::cosine_similarity_batch(query_embedding, vectors, tombstones)
-        }
-    };
+    let vec_scores = exact_vector_scores(query_embedding, vectors, tombstones);
     let kw_scores = bm25_index.scores(query_text);
 
     fuse(vec_scores, kw_scores, fusion, k)
 }
 
+/// Cosine similarity of `query_embedding` to every vector whose `skip` byte is
+/// 0 (a tombstone, or any other exclusion mask). Parallel above 10K vectors
+/// when the `parallel` feature is on.
+pub fn exact_vector_scores(
+    query_embedding: &[f32],
+    vectors: &(impl crate::vector_search::VectorSet + Sync + ?Sized),
+    skip: &[u8],
+) -> Vec<(usize, f32)> {
+    #[cfg(feature = "parallel")]
+    {
+        if vectors.count() > 10_000 {
+            return vector_search::parallel_cosine_batch(
+                query_embedding,
+                vectors,
+                skip,
+                vectors.count(),
+            );
+        }
+    }
+    vector_search::cosine_similarity_batch(query_embedding, vectors, skip)
+}
+
 /// Merge pre-computed vector-similarity and keyword scores into a single ranking.
 ///
 /// Both score sets are independently min-max normalized to [0, 1] and combined
@@ -270,7 +273,7 @@ fn normalize_scores(scores: &[(usize, f32)]) -> Vec<(usize, f32)> {
 pub fn rrf_hybrid_search(
     query_embedding: &[f32],
     query_text: &str,
-    vectors: &[Vec<f32>],
+    vectors: &(impl crate::vector_search::VectorSet + Sync + ?Sized),
     _chunks: &[String],
     tombstones: &[u8],
     bm25_index: &BM25Index,
@@ -282,12 +285,12 @@ pub fn rrf_hybrid_search(
     let mut vec_scores = {
         #[cfg(feature = "parallel")]
         {
-            if vectors.len() > 10_000 {
+            if vectors.count() > 10_000 {
                 vector_search::parallel_cosine_batch(
                     query_embedding,
                     vectors,
                     tombstones,
-                    vectors.len(),
+                    vectors.count(),
                 )
             } else {
                 vector_search::cosine_similarity_batch(query_embedding, vectors, tombstones)
@@ -298,7 +301,7 @@ pub fn rrf_hybrid_search(
             vector_search::cosine_similarity_batch(query_embedding, vectors, tombstones)
         }
     };
-    let mut kw_scores = bm25_index.search(query_text, vectors.len());
+    let mut kw_scores = bm25_index.search(query_text, vectors.count());
 
     // Sort both lists descending so rank 1 = best.
     vec_scores.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap_or(std::cmp::Ordering::Equal));

crates/clawhdf5-agent/src/knowledge.rs

@@ -163,12 +163,13 @@ fn levenshtein(a: &str, b: &str) -> usize {
 /// entities-slice-index map, and an entity-id -> relation-indices map (edges
 /// touching that entity as either source or target).
 ///
-/// Built fresh per traversal call rather than cached on `KnowledgeCache`:
-/// entities/relations are plain `pub` `Vec`s that get pushed to directly
-/// (e.g. `schema.rs`'s load path bypasses `add_entity`/`add_relation`), so a
-/// persistent index would need extra bookkeeping to avoid drifting stale. A
-/// one-off O(V+E) build per call is still a large win over the O(V·E) (BFS)
-/// / O(steps·active·E) (spreading activation) scans it replaces.
+/// Cached on `KnowledgeCache` and checked against a fingerprint of the graph
+/// on every use ([`graph_fingerprint`]). entities/relations are plain `pub`
+/// `Vec`s that get changed directly (e.g. `schema.rs`'s load path bypasses
+/// `add_entity`/`add_relation`), so the cache cannot rely on being told about
+/// changes; the fingerprint notices any of them. Rebuilding it on every
+/// traversal instead made a 2-hop BFS over 1K entities 6.5x slower than the
+/// scan it replaced (24 -> 155 µs; `BENCHMARKS.md`, "Knowledge Graph").
 struct AdjacencyIndex {
     entity_index: HashMap<u64, usize>,
     by_entity: HashMap<u64, Vec<usize>>,
@@ -204,6 +205,45 @@ impl AdjacencyIndex {
     }
 }
 
+/// A hash of everything [`AdjacencyIndex`] depends on — each entity's id and
+/// position, each relation's endpoints and position. One linear pass, no
+/// allocation: far cheaper than building the index, which hashes the same
+/// values into two maps.
+fn graph_fingerprint(entities: &[Entity], relations: &[Relation]) -> u64 {
+    // splitmix64-style mixing; order matters, so positions are covered.
+    fn mix(h: u64, v: u64) -> u64 {
+        let mut z = (h ^ v).wrapping_add(0x9E37_79B9_7F4A_7C15);
+        z = (z ^ (z >> 30)).wrapping_mul(0xBF58_476D_1CE4_E5B9);
+        z = (z ^ (z >> 27)).wrapping_mul(0x94D0_49BB_1331_11EB);
+        z ^ (z >> 31)
+    }
+    let mut h = mix(entities.len() as u64, relations.len() as u64);
+    for e in entities {
+        h = mix(h, e.id);
+    }
+    for r in relations {
+        h = mix(mix(h, r.src), r.tgt);
+    }
+    h
+}
+
+/// The cached [`AdjacencyIndex`] and the fingerprint it was built for.
+/// Cloning a `KnowledgeCache` starts the clone with an empty cache.
+#[derive(Default)]
+struct AdjacencyCache(std::sync::Mutex<Option<(u64, std::sync::Arc<AdjacencyIndex>)>>);
+
+impl Clone for AdjacencyCache {
+    fn clone(&self) -> Self {
+        Self::default()
+    }
+}
+
+impl std::fmt::Debug for AdjacencyCache {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.write_str("AdjacencyCache")
+    }
+}
+
 // ---------------------------------------------------------------------------
 // KnowledgeCache
 // ---------------------------------------------------------------------------
@@ -216,6 +256,7 @@ pub struct KnowledgeCache {
     pub alias_strings: Vec<String>,
     pub alias_entity_ids: Vec<i64>,
     next_entity_id: u64,
+    adjacency: AdjacencyCache,
 }
 
 impl KnowledgeCache {
@@ -226,6 +267,7 @@ impl KnowledgeCache {
             alias_strings: Vec::new(),
             alias_entity_ids: Vec::new(),
             next_entity_id: 0,
+            adjacency: AdjacencyCache::default(),
         }
     }
 
@@ -236,9 +278,29 @@ impl KnowledgeCache {
             alias_strings: Vec::new(),
             alias_entity_ids: Vec::new(),
             next_entity_id: next_id,
+            adjacency: AdjacencyCache::default(),
         }
     }
 
+    /// The adjacency index for the graph as it is now: the cached one if the
+    /// graph's fingerprint still matches, otherwise rebuilt and cached.
+    fn adjacency_index(&self) -> std::sync::Arc<AdjacencyIndex> {
+        let fp = graph_fingerprint(&self.entities, &self.relations);
+        let mut slot = self
+            .adjacency
+            .0
+            .lock()
+            .unwrap_or_else(std::sync::PoisonError::into_inner);
+        if let Some((cached_fp, idx)) = slot.as_ref()
+            && *cached_fp == fp
+        {
+            return idx.clone();
+        }
+        let idx = std::sync::Arc::new(AdjacencyIndex::build(&self.entities, &self.relations));
+        *slot = Some((fp, idx.clone()));
+        idx
+    }
+
     // -----------------------------------------------------------------------
     // Entity management
     // -----------------------------------------------------------------------
@@ -397,7 +459,7 @@ impl KnowledgeCache {
     /// together with their discovered depth.  The seed entity itself is NOT
     /// included.  Traversal follows both outgoing and incoming relation edges.
     pub fn bfs_neighbors(&self, entity_id: u64, max_depth: usize) -> Vec<(Entity, usize)> {
-        let idx = AdjacencyIndex::build(&self.entities, &self.relations);
+        let idx = self.adjacency_index();
         let mut visited: HashSet<u64> = HashSet::new();
         let mut queue: VecDeque<(u64, usize)> = VecDeque::new();
         let mut results: Vec<(Entity, usize)> = Vec::new();
@@ -502,7 +564,7 @@ impl KnowledgeCache {
         min_activation: f32,
         max_steps: usize,
     ) -> Vec<(u64, f32)> {
-        let idx = AdjacencyIndex::build(&self.entities, &self.relations);
+        let idx = self.adjacency_index();
         let mut activation: HashMap<u64, f32> = HashMap::new();
 
         // Initialise seeds with activation 1.0.
@@ -631,6 +693,51 @@ impl Default for KnowledgeCache {
 mod tests {
     use super::*;
 
+    #[test]
+    fn cached_adjacency_sees_direct_changes_to_the_graph() {
+        // The index is cached across traversals, but entities/relations are
+        // pub Vecs anyone can edit; every kind of edit must be seen.
+        let mut kg = KnowledgeCache::new();
+        let a = kg.add_entity("a", "t", -1);
+        let b = kg.add_entity("b", "t", -1);
+        let c = kg.add_entity("c", "t", -1);
+        kg.add_relation(a, b, "r", 1.0);
+        let ids = |kg: &KnowledgeCache| -> Vec<u64> {
+            let mut v: Vec<u64> = kg.bfs_neighbors(a, 3).iter().map(|(e, _)| e.id).collect();
+            v.sort();
+            v
+        };
+        assert_eq!(ids(&kg), vec![b]);
+        assert_eq!(ids(&kg), vec![b], "cached index reused");
+
+        // Pushed directly, bypassing add_relation.
+        kg.relations.push(Relation {
+            src: b,
+            tgt: c,
+            ..Relation::default()
+        });
+        assert_eq!(ids(&kg), vec![b, c]);
+
+        // Rewired in place: same lengths, different edge.
+        kg.relations[1].tgt = a;
+        assert_eq!(ids(&kg), vec![b]);
+
+        // Removed and replaced: same lengths again.
+        kg.relations.pop();
+        kg.relations.push(Relation {
+            src: a,
+            tgt: c,
+            ..Relation::default()
+        });
+        assert_eq!(ids(&kg), vec![b, c]);
+        let act: Vec<u64> = kg
+            .spreading_activation(&[a], 0.5, 0.0, 2)
+            .iter()
+            .map(|(id, _)| *id)
+            .collect();
+        assert!(act.contains(&c));
+    }
+
     // -----------------------------------------------------------------------
     // Original tests — must remain passing
     // -----------------------------------------------------------------------

crates/clawhdf5-agent/src/lib.rs

@@ -1,4 +1,4 @@
-//! ZeroClaw agent memory HDF5 backend.
+//! Agent memory stored in a single HDF5 file.
 //!
 //! Provides persistent memory storage for AI agents using HDF5 files.
 //! All data is cached in-memory for fast access and flushed to disk
@@ -36,6 +36,7 @@ pub mod reranker;
 pub mod schema;
 pub mod search;
 pub mod session;
+pub mod signing;
 pub mod storage;
 mod store_lock;
 pub mod temporal;
@@ -62,26 +63,23 @@ use std::path::{Path, PathBuf};
 
 use cache::MemoryCache;
 #[cfg(feature = "hnsw")]
-use clawhdf5_ann::{DistanceMetric, HnswIndex};
+use clawhdf5_ann::{DistanceMetric, HnswIndex, Storage};
+use clawhdf5_format::float16::round_to_f16;
 use ephemeral::{EphemeralConfig, EphemeralStore};
 
-/// HNSW construction parameters used for the agent's vector index. Cosine is the
-/// agent's similarity metric, so the index is built with cosine distance.
-#[cfg(feature = "hnsw")]
-const HNSW_M: usize = 16;
-#[cfg(feature = "hnsw")]
-const HNSW_EF_CONSTRUCTION: usize = 64;
 // EphemeralEntry and EphemeralStats are part of the crate public API via
 // the `ephemeral` module; they are not needed directly in lib.rs internals.
 #[allow(unused_imports)]
 pub use ephemeral::{EphemeralEntry, EphemeralStats};
 use knowledge::KnowledgeCache;
 use memory_strategy::{Exchange, MemoryStrategy, StrategyOutput};
-use session::SessionCache;
+pub use search::SearchOptions;
+pub use session::{SessionCache, SessionEntry};
 
 // --- Error type ---
 
 #[derive(Debug)]
+#[non_exhaustive]
 pub enum MemoryError {
     Io(std::io::Error),
     Hdf5(String),
@@ -89,6 +87,14 @@ pub enum MemoryError {
     NotFound(String),
     /// Another `HDF5Memory` (in this or another process) has the store open.
     Locked(String),
+    /// A record the store cannot hold as given, e.g. an embedding value
+    /// outside the half-precision range of a `float16` store.
+    InvalidEntry(String),
+    /// The store's checkpoints are signed and no signing key is set, so a
+    /// checkpoint would leave it unsigned. Set the key with
+    /// [`HDF5Memory::set_signing_key`], or drop the signature on purpose with
+    /// [`HDF5Memory::remove_signature`].
+    SigningKeyRequired(String),
 }
 
 impl std::fmt::Display for MemoryError {
@@ -99,6 +105,8 @@ impl std::fmt::Display for MemoryError {
             MemoryError::Schema(e) => write!(f, "schema error: {e}"),
             MemoryError::NotFound(e) => write!(f, "not found: {e}"),
             MemoryError::Locked(e) => write!(f, "store is locked: {e}"),
+            MemoryError::InvalidEntry(e) => write!(f, "invalid entry: {e}"),
+            MemoryError::SigningKeyRequired(e) => write!(f, "signing key required: {e}"),
         }
     }
 }
@@ -130,6 +138,19 @@ pub struct MemoryConfig {
     pub embedding_dim: usize,
     pub chunk_size: usize,
     pub overlap: usize,
+    /// Store embeddings as IEEE half precision (numpy `float16`): half the
+    /// bytes of the embeddings dataset on disk. Every embedding is rounded to
+    /// the nearest half as it enters the store, in memory as well as on disk,
+    /// so search results are the same before and after a reopen. Values must
+    /// lie within ±65504; a save outside that is `MemoryError::InvalidEntry`.
+    /// Fixed when the store is created (persisted in `/meta`).
+    ///
+    /// **On by default for new stores**: on the full LongMemEval haystack with
+    /// real MiniLM embeddings every retrieval metric matched `f32`, and at
+    /// 100K records the file is 48% smaller (`BENCHMARKS.md`). Existing
+    /// stores keep the setting they were created with. Set it to `false` for
+    /// full-precision embeddings, e.g. for unnormalised vectors that may
+    /// exceed the half-precision range.
     pub float16: bool,
     pub compression: bool,
     pub compression_level: u32,
@@ -139,6 +160,40 @@ pub struct MemoryConfig {
     pub created_at: String,
     pub wal_enabled: bool,
     pub wal_max_entries: usize,
+    /// Store the vector index's own copy of the embeddings as int8 rather than
+    /// f32, a quarter of the memory. **On by default** for new stores.
+    ///
+    /// The index's copy is the single largest part of a loaded store's
+    /// footprint. Quantised distances are approximate, so the candidate pool
+    /// is re-scored against the cache's exact embeddings before fusion, which
+    /// holds recall at the f32 index's level. It is also faster, not slower:
+    /// at equal recall, 1.63x the queries per second on x86-64 (AVX2) and
+    /// 1.18x on a Raspberry Pi 5 (NEON `SDOT`), with builds 1.8x and 2.3x
+    /// faster. See `BENCHMARKS.md`.
+    ///
+    /// Persisted with the store. Stores written before this setting existed
+    /// have no stored value and open as `false`, so reopening an old store
+    /// never changes how its index is held.
+    ///
+    /// Has no effect without the `hnsw` feature.
+    pub quantized_index: bool,
+    /// HNSW graph degree. Higher means a denser graph: better recall, more
+    /// memory and slower builds. Clamped to at least 2 when the index is
+    /// built, since a graph with fewer connections is not one.
+    ///
+    /// Has no effect without the `hnsw` feature.
+    pub hnsw_m: usize,
+    /// Candidate list size while building the HNSW graph. Higher means a
+    /// better graph and a slower build; it does not affect query cost.
+    ///
+    /// Has no effect without the `hnsw` feature.
+    pub hnsw_ef_construction: usize,
+    /// Candidate list size for a query, trading throughput for recall. `0`
+    /// keeps the default, which scales with the requested `k`
+    /// (`max(k * 8, 64)`) so that fusion still sees a useful pool.
+    ///
+    /// Has no effect without the `hnsw` feature.
+    pub hnsw_ef_search: usize,
 }
 
 impl MemoryConfig {
@@ -151,7 +206,7 @@ impl MemoryConfig {
             embedding_dim,
             chunk_size: 512,
             overlap: 50,
-            float16: false,
+            float16: true,
             compression: false,
             compression_level: 0,
             compact_threshold: 0.3,
@@ -160,6 +215,10 @@ impl MemoryConfig {
             created_at,
             wal_enabled: true,
             wal_max_entries: 500,
+            quantized_index: true,
+            hnsw_m: 16,
+            hnsw_ef_construction: 64,
+            hnsw_ef_search: 0,
         }
     }
 }
@@ -266,6 +325,12 @@ pub struct HDF5Memory {
     activations_dirty: bool,
     /// Opened with [`HDF5Memory::open_read_only`]: nothing may reach the disk.
     read_only: bool,
+    /// Key that signs every checkpoint; never persisted. See
+    /// [`HDF5Memory::set_signing_key`].
+    signing_key: Option<signing::SigningKey>,
+    /// Checkpoints of this store are signed: the file on disk is, or a key
+    /// has been set. A checkpoint without a key is then refused.
+    signed: bool,
     /// A WAL that `open()` could not read and moved aside; see
     /// [`HDF5Memory::quarantined_wal`].
     quarantined_wal: Option<PathBuf>,
@@ -285,7 +350,8 @@ impl HDF5Memory {
     /// Create a new HDF5 memory file with the given configuration.
     pub fn create(config: MemoryConfig) -> Result<Self> {
         let lock = store_lock::StoreLock::acquire(&config.path)?;
-        let cache = MemoryCache::new(config.embedding_dim);
+        let mut cache = MemoryCache::new(config.embedding_dim);
+        cache.set_half_precision(config.float16);
         let sessions = SessionCache::new();
         let knowledge = KnowledgeCache::new();
 
@@ -320,6 +386,8 @@ impl HDF5Memory {
             bm25_filter: bm25::TokenFilter::default(),
             activations_dirty: false,
             read_only: false,
+            signing_key: None,
+            signed: false,
             quarantined_wal: None,
             _lock: Some(lock),
         })
@@ -444,7 +512,17 @@ impl HDF5Memory {
 
         #[cfg(feature = "hnsw")]
         let loaded_index = if replay_only_appended {
-            Self::load_vector_index(path, checkpoint.ann_generation, &cache, n_checkpoint)
+            Self::load_vector_index(
+                path,
+                checkpoint.ann_generation,
+                &cache,
+                n_checkpoint,
+                if config.quantized_index {
+                    Storage::Int8
+                } else {
+                    Storage::Float32
+                },
+            )
         } else {
             None
         };
@@ -488,6 +566,8 @@ impl HDF5Memory {
             bm25_filter: bm25::TokenFilter::default(),
             activations_dirty: false,
             read_only,
+            signing_key: None,
+            signed: checkpoint.signed,
             quarantined_wal,
             _lock: lock,
         })
@@ -558,6 +638,7 @@ impl HDF5Memory {
         generation: Option<u64>,
         cache: &MemoryCache,
         n_checkpoint: usize,
+        storage: Storage,
     ) -> Option<HnswIndex> {
         let generation = generation?;
         let bytes = std::fs::read(Self::vector_index_path(store)).ok()?;
@@ -565,15 +646,17 @@ impl HDF5Memory {
         if u64::from_le_bytes(stamp.try_into().ok()?) != generation {
             return None;
         }
-        let vectors = cache.embeddings.get(..n_checkpoint)?.to_vec();
-        let mut index = HnswIndex::from_graph_bytes(graph, vectors).ok()?;
+        let vectors: Vec<Vec<f32>> = (0..n_checkpoint)
+            .map(|i| cache.embeddings.get(i).map(<[f32]>::to_vec))
+            .collect::<Option<_>>()?;
+        let mut index = HnswIndex::from_graph_bytes_with(graph, vectors, storage).ok()?;
         if index.dimension() != cache.embedding_dim {
             return None;
         }
         // Records appended since (replayed from the WAL) join incrementally.
         for id in n_checkpoint..cache.embeddings.len() {
             if cache.embeddings[id].len() != index.dimension()
-                || index.insert(cache.embeddings[id].clone()) != id
+                || index.insert(cache.embeddings[id].to_vec()) != id
             {
                 return None;
             }
@@ -645,6 +728,39 @@ impl HDF5Memory {
         }
     }
 
+    /// Sign every checkpoint from now on with `key` (Ed25519). The key is
+    /// never written anywhere; set it again after every `open`. Once a store
+    /// is signed, a checkpoint without the key is refused
+    /// ([`MemoryError::SigningKeyRequired`]) rather than silently leaving it
+    /// unsigned. Setting a different key re-signs the store under that key
+    /// from the next checkpoint; a verifier trusting the old key will then
+    /// reject it, which is the point. Call [`AgentMemory::flush_wal`] to sign
+    /// right away.
+    pub fn set_signing_key(&mut self, key: signing::SigningKey) {
+        self.signing_key = Some(key);
+        self.signed = true;
+    }
+
+    /// Stop signing: the next checkpoint writes the store unsigned. The
+    /// deliberate way out of [`MemoryError::SigningKeyRequired`].
+    pub fn remove_signature(&mut self) {
+        self.signing_key = None;
+        self.signed = false;
+    }
+
+    /// Checkpoints of this store are signed (on disk, or from the next
+    /// checkpoint because a key has been set).
+    pub fn is_signed(&self) -> bool {
+        self.signed
+    }
+
+    /// Check the checkpoint at `path` against the public key the caller
+    /// trusts; see [`signing::verify_store`]. Reads the file only: it works
+    /// on a store another process has open.
+    pub fn verify(path: &Path, trusted: &signing::VerifyingKey) -> Result<signing::VerifyReport> {
+        signing::verify_store(path, trusted)
+    }
+
     /// Flush current state to disk and truncate the WAL.
     ///
     /// Every code path that persists the full cache to the .h5 file must
@@ -660,10 +776,28 @@ impl HDF5Memory {
         // Record which WAL prefix this checkpoint contains, so a crash before
         // the truncate below can't replay those entries a second time.
         let wal_applied = self.wal.as_ref().map(|w| w.mark());
+        let signature = match &self.signing_key {
+            Some(key) => Some(signing::sign(
+                key,
+                &self.config,
+                &self.cache,
+                &self.sessions,
+                &self.knowledge,
+                wal_applied,
+            )),
+            None if self.signed => {
+                return Err(MemoryError::SigningKeyRequired(format!(
+                    "{} is signed; set its signing key before a checkpoint \
+                     (saves so far are held in the WAL or in memory)",
+                    self.config.path.display()
+                )));
+            }
+            None => None,
+        };
         // Written before the .h5 so a crash in between leaves a sidecar whose
         // generation matches no checkpoint (ignored), never the reverse.
         let ann_generation = self.persist_vector_index();
-        storage::write_to_disk_with_meta(
+        storage::write_to_disk_signed(
             &self.config.path,
             &self.config,
             &self.cache,
@@ -672,7 +806,9 @@ impl HDF5Memory {
             &schema::CheckpointMeta {
                 wal_applied,
                 ann_generation,
+                signed: signature.is_some(),
             },
+            signature.as_ref(),
         )?;
         if let Some(ref mut w) = self.wal {
             w.truncate()?;
@@ -801,6 +937,42 @@ impl HDF5Memory {
     // the index length drifts from the cache length (covering any mutation path
     // that doesn't call a hook, e.g. consolidation pushes).
 
+    /// Graph degree for the index, never below the 2 the builder requires:
+    /// a config value of 0 or 1 would otherwise panic inside `clawhdf5-ann`.
+    #[cfg(feature = "hnsw")]
+    fn hnsw_m(&self) -> usize {
+        self.config.hnsw_m.max(2)
+    }
+
+    /// Build-time candidate list size, never below the graph degree — a
+    /// smaller one cannot fill a node's connections.
+    #[cfg(feature = "hnsw")]
+    fn hnsw_ef_construction(&self) -> usize {
+        self.config.hnsw_ef_construction.max(self.hnsw_m())
+    }
+
+    /// Query-time candidate list size for a `k`-result search. `0` means the
+    /// default, which scales with `k`.
+    #[cfg(feature = "hnsw")]
+    pub(crate) fn hnsw_ef_search(&self, k: usize) -> usize {
+        let default = (k * 8).max(64);
+        if self.config.hnsw_ef_search == 0 {
+            default
+        } else {
+            self.config.hnsw_ef_search.max(k)
+        }
+    }
+
+    /// How the index should store its copy of the vectors, per the config.
+    #[cfg(feature = "hnsw")]
+    fn index_storage(&self) -> Storage {
+        if self.config.quantized_index {
+            Storage::Int8
+        } else {
+            Storage::Float32
+        }
+    }
+
     /// Build an HNSW index over the entire cache, re-applying tombstones as
     /// soft-deletions so node ids stay aligned with cache indices.
     ///
@@ -816,11 +988,15 @@ impl HDF5Memory {
         if self.cache.embeddings.iter().any(|e| e.len() != dim) {
             return None;
         }
-        let mut index = HnswIndex::build_with_metric(
-            &self.cache.embeddings,
-            HNSW_M,
-            HNSW_EF_CONSTRUCTION,
+        // The index owns its vectors, so it needs rows rather than the cache's
+        // flat buffer. This copy is the index's own; the cache keeps one.
+        let rows: Vec<Vec<f32>> = self.cache.embeddings.iter().map(<[f32]>::to_vec).collect();
+        let mut index = HnswIndex::build_with(
+            &rows,
+            self.hnsw_m(),
+            self.hnsw_ef_construction(),
             DistanceMetric::Cosine,
+            self.index_storage(),
         );
         for (i, &t) in self.cache.tombstones.iter().enumerate() {
             if t != 0 {
@@ -846,7 +1022,7 @@ impl HDF5Memory {
             let dim = index.dimension();
             let appended = (self.hnsw_synced_len..n).all(|id| {
                 self.cache.embeddings[id].len() == dim
-                    && index.insert(self.cache.embeddings[id].clone()) == id
+                    && index.insert(self.cache.embeddings[id].to_vec()) == id
             });
             if appended {
                 for id in self.hnsw_synced_len..n {
@@ -877,7 +1053,7 @@ impl HDF5Memory {
         let emb_len = self.cache.embeddings[idx].len();
         match self.hnsw.as_mut() {
             Some(index) if emb_len == index.dimension() => {
-                let id = index.insert(self.cache.embeddings[idx].clone());
+                let id = index.insert(self.cache.embeddings[idx].to_vec());
                 if id == idx {
                     self.hnsw_synced_len = self.cache.embeddings.len();
                 } else {
@@ -923,6 +1099,18 @@ impl HDF5Memory {
         &self.config
     }
 
+    /// The sessions recorded in this store.
+    pub fn sessions(&self) -> &SessionCache {
+        &self.sessions
+    }
+
+    /// Mutable access to the sessions, e.g. to add many at once. Changes
+    /// reach the disk at the next checkpoint (any flushing call, such as
+    /// [`HDF5Memory::flush_wal`] or `save_batch`), not immediately.
+    pub fn sessions_mut(&mut self) -> &mut SessionCache {
+        &mut self.sessions
+    }
+
     /// Get a reference to the knowledge cache.
     pub fn knowledge(&self) -> &KnowledgeCache {
         &self.knowledge
@@ -985,7 +1173,29 @@ impl HDF5Memory {
     /// Upsert: if an active entry with the same tags (key) exists, update it in-place.
     /// Otherwise append a new entry. Use this for key-based memory stores where
     /// the same key should not create duplicates.
+    /// A `float16` store holds embeddings as IEEE half precision, which has no
+    /// finite value beyond ±65504. Refuse such an embedding rather than
+    /// silently store infinity. (Values that are already infinite or NaN are
+    /// stored as they are, as in an `f32` store.)
+    fn check_embedding(&self, embedding: &[f32]) -> Result<()> {
+        if !self.config.float16 {
+            return Ok(());
+        }
+        let overflow = embedding
+            .iter()
+            .enumerate()
+            .find(|&(_, &v)| v.is_finite() && round_to_f16(v).is_infinite());
+        match overflow {
+            None => Ok(()),
+            Some((i, v)) => Err(MemoryError::InvalidEntry(format!(
+                "embedding[{i}] = {v} is outside the half-precision range (±65504) \
+                 of this float16 store"
+            ))),
+        }
+    }
+
     pub fn save_or_update(&mut self, entry: MemoryEntry) -> Result<usize> {
+        self.check_embedding(&entry.embedding)?;
         if let Some(existing_idx) = self.cache.find_by_tags(&entry.tags) {
             if let Some(ref mut w) = self.wal {
                 let wal_entry = wal::WalEntry {
@@ -1041,6 +1251,7 @@ impl HDF5Memory {
 
 impl AgentMemory for HDF5Memory {
     fn save(&mut self, entry: MemoryEntry) -> Result<usize> {
+        self.check_embedding(&entry.embedding)?;
         if let Some(ref mut w) = self.wal {
             let wal_entry = wal::WalEntry {
                 entry_type: wal::WalEntryType::Save,
@@ -1082,6 +1293,10 @@ impl AgentMemory for HDF5Memory {
     }
 
     fn save_batch(&mut self, entries: Vec<MemoryEntry>) -> Result<Vec<usize>> {
+        // All or nothing: check every entry before storing any.
+        for entry in &entries {
+            self.check_embedding(&entry.embedding)?;
+        }
         let mut indices = Vec::with_capacity(entries.len());
         for entry in entries {
             let idx = self.cache.push(
@@ -1242,6 +1457,9 @@ impl HDF5Memory {
         })?;
         let view = memory_strategy::CacheStoreView::new(&self.cache, &self.knowledge);
         let output = strat.evaluate(&exchange, &view);
+        for e in &output.entries {
+            self.check_embedding(&e.embedding)?;
+        }
         for e in &output.entries {
             self.cache.push(
                 e.chunk.clone(),
@@ -1266,6 +1484,35 @@ impl HDF5Memory {
 }
 
 impl HDF5Memory {
+    /// Delete many records with a single checkpoint, where
+    /// [`AgentMemory::delete`] checkpoints once per record.
+    ///
+    /// All or nothing: if any id is out of range or already deleted (or
+    /// repeated), nothing is deleted and `MemoryError::NotFound` is returned.
+    /// Unlike `delete`, this never auto-compacts, so the records stay in the
+    /// store as tombstones (their indices unchanged) until [`AgentMemory::compact`]
+    /// is called — importers use it to carry over records that were already
+    /// deleted in the source.
+    pub fn delete_batch(&mut self, ids: &[usize]) -> Result<()> {
+        let mut seen = std::collections::HashSet::with_capacity(ids.len());
+        for &id in ids {
+            if self.cache.tombstones.get(id).copied() != Some(0) || !seen.insert(id) {
+                return Err(MemoryError::NotFound(format!(
+                    "entry {id} not found or already deleted"
+                )));
+            }
+        }
+        if ids.is_empty() {
+            return Ok(());
+        }
+        for &id in ids {
+            self.cache.mark_deleted(id);
+            self.hnsw_on_delete(id);
+            self.bm25_on_delete(id);
+        }
+        self.flush()
+    }
+
     pub fn tick_session(&mut self) -> Result<()> {
         let d = self.config.decay_factor;
         for w in self.cache.activation_weights.iter_mut() {
@@ -1326,6 +1573,16 @@ impl HDF5Memory {
         let mut promoted = 0;
 
         for key in candidates {
+            // Check before taking, so a rejected entry stays in the ephemeral
+            // tier rather than being lost.
+            if let Some(emb) = self
+                .ephemeral
+                .as_ref()
+                .and_then(|s| s.get_entry(&key))
+                .and_then(|e| e.embedding.as_deref())
+            {
+                self.check_embedding(emb)?;
+            }
             let entry = match self
                 .ephemeral
                 .as_mut()
@@ -1454,6 +1711,79 @@ mod tests {
         }
     }
 
+    #[test]
+    fn delete_batch_tombstones_without_compacting() {
+        let dir = TempDir::new().unwrap();
+        let path = dir.path().join("test.h5");
+        let mut mem = HDF5Memory::create(make_config(&dir)).unwrap();
+        mem.save_batch(
+            (0..4)
+                .map(|i| make_entry(&format!("record {i}"), &[i as f32, 1.0, 0.0, 0.0]))
+                .collect(),
+        )
+        .unwrap();
+        // 3 of 4 is far past compact_threshold (0.3): delete() would compact.
+        mem.delete_batch(&[0, 1, 3]).unwrap();
+        assert_eq!(mem.count(), 4);
+        assert_eq!(mem.count_active(), 1);
+        drop(mem);
+
+        let mut mem = HDF5Memory::open(&path).unwrap();
+        assert_eq!(mem.cache.tombstones, vec![1, 1, 0, 1]);
+        let hits = mem.hybrid_search(&[0.0, 1.0, 0.0, 0.0], "record", 0.5, 0.5, 10);
+        assert!(
+            hits.iter().all(|r| r.index == 2),
+            "tombstoned record returned"
+        );
+    }
+
+    #[test]
+    fn delete_batch_is_all_or_nothing() {
+        let dir = TempDir::new().unwrap();
+        let mut mem = HDF5Memory::create(make_config(&dir)).unwrap();
+        mem.save_batch(vec![
+            make_entry("a", &[1.0, 0.0, 0.0, 0.0]),
+            make_entry("b", &[0.0, 1.0, 0.0, 0.0]),
+        ])
+        .unwrap();
+        for bad in [&[0, 5][..], &[1, 1][..]] {
+            assert!(matches!(
+                mem.delete_batch(bad),
+                Err(MemoryError::NotFound(_))
+            ));
+            assert_eq!(mem.count_active(), 2, "{bad:?} deleted something");
+        }
+        mem.delete_batch(&[]).unwrap();
+        assert_eq!(mem.count_active(), 2);
+    }
+
+    #[test]
+    fn sessions_mut_add_at_keeps_timestamp_across_reopen() {
+        let dir = TempDir::new().unwrap();
+        let path = dir.path().join("test.h5");
+        let mut mem = HDF5Memory::create(make_config(&dir)).unwrap();
+        mem.sessions_mut()
+            .add_at("s-old", 2, 7, "discord", "old summary", 1.7e15);
+        mem.flush_wal().unwrap();
+        drop(mem);
+
+        let mem = HDF5Memory::open_read_only(&path).unwrap();
+        let s = mem.sessions();
+        assert_eq!(s.len(), 1);
+        let e = &s.entries[0];
+        assert_eq!(
+            (
+                e.id.as_str(),
+                e.start_idx,
+                e.end_idx,
+                e.channel.as_str(),
+                e.ts
+            ),
+            ("s-old", 2, 7, "discord", 1.7e15)
+        );
+        assert_eq!(s.summaries[0], "old summary");
+    }
+
     #[test]
     fn create_new_file() {
         let dir = TempDir::new().unwrap();

crates/clawhdf5-agent/src/openclaw.rs

@@ -1,10 +1,12 @@
-//! OpenClaw Integration Layer.
+//! A Markdown-oriented memory backend over [`crate::HDF5Memory`].
 //!
-//! Bridge between OpenClaw agent gateway (Markdown + sqlite-vec) and the
-//! clawhdf5 HDF5-backed memory backend.  Provides:
+//! Named for OpenClaw, whose workspace memory is Markdown, but **not an
+//! OpenClaw plugin**: nothing here registers with OpenClaw, and the
+//! integration it was written for never worked (see `docs/openclaw.md`).
+//! Provides:
 //!
-//! - [`MemoryBackend`] — the trait OpenClaw implements against.
-//! - [`ClawhdfBackend`] — concrete HDF5-backed implementation.
+//! - [`MemoryBackend`] — search / read back / write / ingest / export.
+//! - [`ClawhdfBackend`] — the HDF5-backed implementation.
 //! - [`MarkdownParser`] — splits Markdown into [`MarkdownSection`] records.
 //! - [`MarkdownExporter`] — renders sections back to Markdown text.
 
@@ -13,9 +15,8 @@ use std::path::{Path, PathBuf};
 use std::time::{SystemTime, UNIX_EPOCH};
 
 use crate::{
-    AgentMemory, HDF5Memory, MemoryConfig, MemoryEntry,
-    confidence::{ConfidenceConfig, ScoredResult, reject_low_confidence},
-    reranker::{ReRankConfig, RerankInput, rerank},
+    AgentMemory, HDF5Memory, MemoryConfig, MemoryEntry, SearchOptions,
+    confidence::ConfidenceConfig, reranker::ReRankConfig,
 };
 
 // ─────────────────────────────────────────────────────────────────────────────
@@ -62,7 +63,8 @@ pub struct BackendStats {
 // MemoryBackend trait
 // ─────────────────────────────────────────────────────────────────────────────
 
-/// Interface that OpenClaw uses to interact with a memory backend.
+/// A Markdown-oriented memory backend: search, read back by path, write,
+/// ingest and export.
 ///
 /// Implementors provide persistent storage, full-text + vector search,
 /// Markdown ingestion / export, and statistics.
@@ -319,7 +321,7 @@ impl MarkdownExporter {
 ///
 /// # Path mapping
 ///
-/// OpenClaw addresses memories by file path (e.g. `"memory/user.md"`).
+/// Memories are addressed by file path (e.g. `"memory/user.md"`).
 /// Internally every [`MemoryEntry`] stores the originating path as its
 /// `source_channel`.  Section sub-paths are stored as
 /// `"<path>::<heading>"`.
@@ -422,7 +424,7 @@ impl ClawhdfBackend {
 
     // ── Compaction & Consolidation hooks (7.6) ────────────────────────────
 
-    /// Run a compaction cycle — called by OpenClaw during session compaction.
+    /// Run a compaction cycle (decay, compaction, WAL flush).
     ///
     /// Sequence:
     /// 1. `tick_session()` — apply Hebbian decay to all activation weights.
@@ -466,7 +468,7 @@ impl ClawhdfBackend {
             let record = MemoryRecord {
                 id: i as u64,
                 chunk: cache.chunks[i].clone(),
-                embedding: cache.embeddings[i].clone(),
+                embedding: cache.embeddings[i].to_vec(),
                 tier: MemoryTier::Working,
                 importance: cache.activation_weights[i],
                 access_count: 0,
@@ -524,70 +526,27 @@ impl ClawhdfBackend {
 
 impl MemoryBackend for ClawhdfBackend {
     /// Search using hybrid vector + BM25 retrieval, then re-rank and
-    /// confidence-filter.
+    /// confidence-filter — [`HDF5Memory::search`] with both stages on.
     fn search(
         &mut self,
         query_text: &str,
         query_embedding: &[f32],
         k: usize,
     ) -> Vec<MemorySearchResult> {
-        // 1. Hybrid retrieval (vector + BM25, fused by score).
-        let candidates = k.saturating_mul(3).max(10);
-        let raw = self.memory.hybrid_search_with(
-            query_embedding,
-            query_text,
-            crate::hybrid::DEFAULT_FUSION,
-            candidates,
-        );
-
-        if raw.is_empty() {
-            return Vec::new();
-        }
-
-        let now = Self::now_secs();
-
-        // 2. Re-rank using temporal recency, source authority, Hebbian weight.
-        let rerank_inputs: Vec<RerankInput> = raw
-            .iter()
-            .map(|r| RerankInput {
-                index: r.index,
-                timestamp: r.timestamp,
-                source_channel: r.source_channel.clone(),
-                raw_activation: r.activation,
-            })
-            .collect();
-
-        let reranked = rerank(&rerank_inputs, &self.rerank_config, now);
-
-        // 3. Confidence rejection.
-        let scored: Vec<ScoredResult> = reranked
-            .iter()
-            .map(|r| ScoredResult {
-                index: r.index,
-                score: r.combined_score,
-            })
-            .collect();
-
-        let confident = reject_low_confidence(&scored, &self.confidence_config);
-
-        // 4. Map back to MemorySearchResult; preserve raw text via index lookup.
-        let raw_by_idx: HashMap<usize, &crate::SearchResult> =
-            raw.iter().map(|r| (r.index, r)).collect();
-
-        confident
+        let options = SearchOptions::new(k)
+            .with_rerank(self.rerank_config)
+            .with_confidence(self.confidence_config.clone())
+            .at_time(Self::now_secs());
+        self.memory
+            .search(query_embedding, query_text, &options)
             .into_iter()
-            .take(k)
-            .filter_map(|sr| {
-                let r = raw_by_idx.get(&sr.index)?;
-                let path = r.source_channel.clone();
-                Some(MemorySearchResult {
-                    text: r.chunk.clone(),
-                    score: sr.score,
-                    path: path.clone(),
+            .map(|r| MemorySearchResult {
+                text: r.chunk,
+                score: r.score,
+                path: r.source_channel.clone(),
                 line_range: None,
                 timestamp: Some(r.timestamp),
-                    source: path,
-                })
+                source: r.source_channel,
             })
             .collect()
     }
@@ -716,11 +675,13 @@ impl MemoryBackend for ClawhdfBackend {
 
         let total_records = cache.count_active();
 
+        // A record saved without an embedding occupies a zero row, so "has an
+        // embedding" is "has a non-zero norm" rather than "row is non-empty".
         let total_embeddings = cache
-            .embeddings
+            .norms
             .iter()
             .enumerate()
-            .filter(|(i, emb)| cache.tombstones[*i] == 0 && !emb.is_empty())
+            .filter(|(i, norm)| cache.tombstones[*i] == 0 && **norm > 0.0)
             .count();
 
         let file_size_bytes = std::fs::metadata(&self.hdf5_path)

crates/clawhdf5-agent/src/provenance.rs

@@ -2,7 +2,7 @@
 //!
 //! Records the origin, authorship, and a content hash of every memory chunk
 //! so the system can detect *accidental* corruption and trace data lineage.
-//! The hash is unkeyed (see [`fnv1a_64`]) — this is not a tamper-evidence or
+//! The hash is unkeyed (FNV-1a) — this is not a tamper-evidence or
 //! authenticity guarantee.
 
 use std::collections::HashMap;

crates/clawhdf5-agent/src/reranker.rs

@@ -4,8 +4,10 @@
 //! into a single composite score for each retrieved result.
 
 /// Configuration for the multi-factor re-ranker.
-#[derive(Debug, Clone)]
+#[derive(Debug, Clone, Copy)]
 pub struct ReRankConfig {
+    /// Weight applied to the retrieval score the candidate arrived with.
+    pub relevance_weight: f32,
     /// Weight applied to the temporal decay score (0.0–1.0).
     pub temporal_weight: f32,
     /// Weight applied to the source authority score (0.0–1.0).
@@ -20,6 +22,9 @@ pub struct ReRankConfig {
 impl Default for ReRankConfig {
     fn default() -> Self {
         Self {
+            // Relevance leads: the metadata signals break ties and nudge, they
+            // do not decide. See `BENCHMARKS.md`, "Recency discrimination".
+            relevance_weight: 1.0,
             temporal_weight: 0.3,
             authority_weight: 0.2,
             activation_weight: 0.5,
@@ -41,6 +46,8 @@ pub struct ReRankResult {
     pub authority_score: f32,
     /// Normalised Hebbian activation score in [0, 1].
     pub activation_score: f32,
+    /// The retrieval score carried through from the input.
+    pub relevance_score: f32,
 }
 
 /// Compute an exponential decay temporal score.
@@ -105,6 +112,15 @@ pub struct RerankInput {
     pub source_channel: String,
     /// Raw Hebbian activation weight for this entry.
     pub raw_activation: f32,
+    /// The retrieval score that put this entry in the candidate list.
+    ///
+    /// Re-ranking is meant to *adjust* the retriever's ordering with signals
+    /// it does not have, not to replace it. Without this the combined score
+    /// was made of recency, authority and activation alone, so a candidate
+    /// pool came back ordered by age with its relevance ordering discarded.
+    /// Callers with no meaningful score can pass the same value for every
+    /// entry, which reduces to the old behaviour.
+    pub relevance: f32,
 }
 
 /// Re-rank a list of retrieval results using multi-factor scoring.
@@ -138,7 +154,8 @@ pub fn rerank(
             let auth = source_authority_score(&inp.source_channel);
             let act = activation_score(inp.raw_activation);
 
-            let combined = config.temporal_weight * ts
+            let combined = config.relevance_weight * inp.relevance
+                + config.temporal_weight * ts
                 + config.authority_weight * auth
                 + config.activation_weight * act;
 
@@ -148,6 +165,7 @@ pub fn rerank(
                 temporal_score: ts,
                 authority_score: auth,
                 activation_score: act,
+                relevance_score: inp.relevance,
             }
         })
         .collect();
@@ -253,22 +271,51 @@ mod tests {
                 timestamp: 0.0, // very old
                 source_channel: "other".to_string(),
                 raw_activation: 0.1,
+                relevance: 0.0,
             },
             RerankInput {
                 index: 1,
                 timestamp: 86_400.0, // one day ago
                 source_channel: "conversation".to_string(),
                 raw_activation: 0.5,
+                relevance: 0.0,
             },
             RerankInput {
                 index: 2,
                 timestamp: 172_800.0, // "now"
                 source_channel: "user_correction".to_string(),
                 raw_activation: 1.0,
+                relevance: 0.0,
             },
         ]
     }
 
+    #[test]
+    fn relevance_leads_but_recency_breaks_near_ties() {
+        let entry = |index, timestamp, relevance| RerankInput {
+            index,
+            timestamp,
+            source_channel: "conversation".to_string(),
+            raw_activation: 1.0,
+            relevance,
+        };
+        let now = 10.0 * 86_400.0;
+        let config = ReRankConfig::default();
+
+        // A clearly better match wins despite being much older. Before
+        // `relevance` existed the combined score ignored it entirely, so this
+        // returned the newer, irrelevant entry.
+        let ranked = rerank(&[entry(0, 0.0, 1.0), entry(1, now, 0.1)], &config, now);
+        assert_eq!(ranked[0].index, 0, "{ranked:?}");
+
+        // Between near-equal matches, the newer one wins.
+        let ranked = rerank(&[entry(0, 0.0, 0.80), entry(1, now, 0.79)], &config, now);
+        assert_eq!(ranked[0].index, 1, "{ranked:?}");
+
+        // The breakdown carries the relevance through.
+        assert_eq!(ranked[0].relevance_score, 0.79);
+    }
+
     #[test]
     fn rerank_returns_all_entries() {
         let inputs = make_inputs();
@@ -302,6 +349,7 @@ mod tests {
     #[test]
     fn rerank_score_breakdown_matches_manual_calculation() {
         let config = ReRankConfig {
+            relevance_weight: 0.0,
             temporal_weight: 1.0,
             authority_weight: 0.0,
             activation_weight: 0.0,
@@ -312,6 +360,7 @@ mod tests {
             timestamp: 0.0,
             source_channel: "other".to_string(),
             raw_activation: 0.5,
+            relevance: 0.0,
         }];
         let now = 3600.0_f64; // exactly one half-life later
         let results = rerank(&inputs, &config, now);

crates/clawhdf5-agent/src/schema.rs

@@ -15,6 +15,9 @@ use crate::session::SessionCache;
 use crate::wal::WalMark;
 
 pub const SCHEMA_VERSION: &str = "1.0";
+/// Writer-version tag stored in `/meta` as `edgehdf5_version`. Kept for file
+/// compatibility; despite the name it has nothing to do with ZeroClaw, which
+/// does not use clawhdf5.
 pub const ZEROCLAW_VERSION: &str = "0.8.0";
 
 /// `/meta` attributes holding the [`WalMark`] of the WAL prefix already folded
@@ -23,6 +26,7 @@ pub const ZEROCLAW_VERSION: &str = "0.8.0";
 const WAL_APPLIED_LEN_ATTR: &str = "wal_applied_len";
 const WAL_APPLIED_CRC_ATTR: &str = "wal_applied_crc";
 const ANN_GENERATION_ATTR: &str = "ann_generation";
+const SIG_VERSION_ATTR: &str = "sig_version";
 
 /// Build a complete HDF5 file from the in-memory state.
 pub fn build_hdf5_file(
@@ -46,7 +50,7 @@ pub fn build_hdf5_file_with_mark(
 ) -> Result<Vec<u8>, MemoryError> {
     let meta = CheckpointMeta {
         wal_applied,
-        ann_generation: None,
+        ..CheckpointMeta::default()
     };
     build_hdf5_file_with_meta(config, cache, sessions, knowledge, &meta)
 }
@@ -61,6 +65,10 @@ pub struct CheckpointMeta {
     /// one left over from another checkpoint can never be attached to records
     /// it wasn't built from.
     pub ann_generation: Option<u64>,
+    /// The checkpoint carries an Ed25519 signature (see [`crate::signing`]).
+    /// Read-only: whether a checkpoint is *written* signed is decided by the
+    /// signature passed to [`build_hdf5_file_signed`].
+    pub signed: bool,
 }
 
 /// [`build_hdf5_file`] with checkpoint bookkeeping.
@@ -70,6 +78,19 @@ pub fn build_hdf5_file_with_meta(
     sessions: &SessionCache,
     knowledge: &KnowledgeCache,
     checkpoint: &CheckpointMeta,
+) -> Result<Vec<u8>, MemoryError> {
+    build_hdf5_file_signed(config, cache, sessions, knowledge, checkpoint, None)
+}
+
+/// [`build_hdf5_file_with_meta`], plus a signed manifest of the contents
+/// (see [`crate::signing`]).
+pub fn build_hdf5_file_signed(
+    config: &MemoryConfig,
+    cache: &MemoryCache,
+    sessions: &SessionCache,
+    knowledge: &KnowledgeCache,
+    checkpoint: &CheckpointMeta,
+    signature: Option<&crate::signing::StoredSignature>,
 ) -> Result<Vec<u8>, MemoryError> {
     let wal_applied = checkpoint.wal_applied;
     let mut builder = clawhdf5::FileBuilder::new();
@@ -104,6 +125,19 @@ pub fn build_hdf5_file_with_meta(
         "wal_max_entries",
         AttrValue::I64(config.wal_max_entries as i64),
     );
+    meta.set_attr(
+        "quantized_index",
+        AttrValue::I64(config.quantized_index.into()),
+    );
+    meta.set_attr("hnsw_m", AttrValue::I64(config.hnsw_m as i64));
+    meta.set_attr(
+        "hnsw_ef_construction",
+        AttrValue::I64(config.hnsw_ef_construction as i64),
+    );
+    meta.set_attr(
+        "hnsw_ef_search",
+        AttrValue::I64(config.hnsw_ef_search as i64),
+    );
     meta.set_attr(
         "edgehdf5_version",
         AttrValue::String(ZEROCLAW_VERSION.into()),
@@ -117,11 +151,42 @@ pub fn build_hdf5_file_with_meta(
         // round trip through every reader.
         meta.set_attr(ANN_GENERATION_ATTR, AttrValue::I64(generation as i64));
     }
+    if let Some(sig) = signature {
+        use crate::signing::to_hex;
+        let m = &sig.manifest;
+        meta.set_attr(
+            SIG_VERSION_ATTR,
+            AttrValue::I64(crate::signing::MANIFEST_VERSION),
+        );
+        meta.set_attr("sig_algorithm", AttrValue::String("ed25519".into()));
+        meta.set_attr("sig_public_key", AttrValue::String(to_hex(&sig.public_key)));
+        meta.set_attr("sig_signature", AttrValue::String(to_hex(&sig.signature)));
+        meta.set_attr("sig_record_count", AttrValue::I64(m.record_count as i64));
+        meta.set_attr(
+            "sig_records_root",
+            AttrValue::String(to_hex(&m.records_root)),
+        );
+        meta.set_attr("sig_settings", AttrValue::String(to_hex(&m.settings)));
+        meta.set_attr("sig_sessions", AttrValue::String(to_hex(&m.sessions)));
+        meta.set_attr("sig_graph", AttrValue::String(to_hex(&m.graph)));
+    }
     // Need at least one dataset in the group for it to be a proper group
     meta.create_dataset("_marker").with_u8_data(&[1]).compact();
     let finished_meta = meta.finish();
     builder.add_group(finished_meta);
 
+    // /integrity: the signed per-record hashes, so verification can say
+    // which records changed.
+    if let Some(sig) = signature {
+        let mut group = builder.create_group("integrity");
+        let flat: Vec<u8> = sig.record_hashes.iter().flatten().copied().collect();
+        group
+            .create_dataset("record_hashes")
+            .with_u8_data(&flat)
+            .with_shape(&[sig.record_hashes.len() as u64, 32]);
+        builder.add_group(group.finish());
+    }
+
     // /memory group
     build_memory_group(&mut builder, config, cache)?;
 
@@ -146,20 +211,27 @@ fn build_memory_group(
     // chunks: fixed-length string array
     write_string_dataset(&mut group, "chunks", &cache.chunks);
 
-    // embeddings: f32 [N x D]
+    // embeddings: [N x D], f32 — or IEEE half precision for a `float16`
+    // store. The cache already holds half-rounded values then, so this
+    // conversion is exact and a reopened store sees the same numbers.
     let n = cache.embeddings.len() as u64;
     let d = cache.embedding_dim as u64;
     let flat = cache.flat_embeddings();
     {
-        let ds = group
-            .create_dataset("embeddings")
-            .with_f32_data(&flat)
-            .with_shape(&[n, d]);
+        let ds = group.create_dataset("embeddings");
+        let elem_bytes: u64 = if config.float16 {
+            ds.with_f16_data(flat);
+            2
+        } else {
+            ds.with_f32_data(flat);
+            4
+        };
+        ds.with_shape(&[n, d]);
 
         // Chunk size tuning: target ~256KB per chunk for optimal I/O
         if n > 0 && d > 0 {
             let target_chunk_bytes: u64 = 256 * 1024;
-            let rows_per_chunk = (target_chunk_bytes / (d * 4)).max(1).min(n);
+            let rows_per_chunk = (target_chunk_bytes / (d * elem_bytes)).max(1).min(n);
             ds.with_chunks(&[rows_per_chunk, d]);
 
             // Compression. Shuffle is applied automatically (auto-shuffle
@@ -405,10 +477,34 @@ fn write_string_dataset(
     }
 }
 
+/// `/meta`'s attributes, failing if any of them cannot be read.
+///
+/// `Group::attrs` leaves out an attribute it cannot decode. For the store's
+/// settings that would silently fall back to defaults (e.g. `float16`, the
+/// WAL mark), so an unreadable attribute is an error here, as it was before
+/// `attrs` became tolerant.
+fn meta_attrs(
+    file: &clawhdf5::File,
+) -> Result<std::collections::HashMap<String, AttrValue>, MemoryError> {
+    let meta = file
+        .group("meta")
+        .map_err(|e| MemoryError::Schema(format!("missing /meta group: {e}")))?;
+    let (attrs, errors) = meta
+        .attrs_with_errors()
+        .map_err(|e| MemoryError::Schema(format!("cannot read /meta attrs: {e}")))?;
+    if let Some(e) = errors.first() {
+        return Err(MemoryError::Schema(format!(
+            "cannot read /meta attrs: {} unreadable, first: {e}",
+            errors.len()
+        )));
+    }
+    Ok(attrs)
+}
+
 /// Validate an HDF5 file has the correct schema and load all data.
 /// Read the checkpoint's [`WalMark`] from `/meta`, if it has one.
 pub fn read_wal_mark(file: &clawhdf5::File) -> Option<WalMark> {
-    let attrs = file.group("meta").ok()?.attrs().ok()?;
+    let attrs = meta_attrs(file).ok()?;
     let len = match attrs.get(WAL_APPLIED_LEN_ATTR)? {
         AttrValue::I64(v) => u64::try_from(*v).ok()?,
         _ => return None,
@@ -420,19 +516,75 @@ pub fn read_wal_mark(file: &clawhdf5::File) -> Option<WalMark> {
     Some(WalMark { len, crc })
 }
 
+/// Read a checkpoint's signature, if it has one. A signature whose
+/// attributes are present but malformed is an error, not "unsigned".
+pub fn read_signature(
+    file: &clawhdf5::File,
+) -> Result<Option<crate::signing::StoredSignature>, MemoryError> {
+    use crate::signing::{Manifest, StoredSignature, from_hex};
+    let attrs = meta_attrs(file)?;
+    let version = match attrs.get(SIG_VERSION_ATTR) {
+        None => return Ok(None),
+        Some(AttrValue::I64(v)) => *v,
+        Some(_) => return Err(MemoryError::Schema("malformed sig_version".into())),
+    };
+    if version != crate::signing::MANIFEST_VERSION {
+        return Err(MemoryError::Schema(format!(
+            "unsupported signature version {version}"
+        )));
+    }
+    fn hex<const N: usize>(
+        attrs: &std::collections::HashMap<String, AttrValue>,
+        name: &str,
+    ) -> Result<[u8; N], MemoryError> {
+        match attrs.get(name) {
+            Some(AttrValue::String(s)) => from_hex::<N>(s),
+            _ => None,
+        }
+        .ok_or_else(|| MemoryError::Schema(format!("malformed or missing {name}")))
+    }
+    let record_count = match attrs.get("sig_record_count") {
+        Some(AttrValue::I64(v)) if *v >= 0 => *v as u64,
+        _ => return Err(MemoryError::Schema("malformed sig_record_count".into())),
+    };
+    let group = file
+        .group("integrity")
+        .map_err(|e| MemoryError::Schema(format!("signed checkpoint without /integrity: {e}")))?;
+    let flat = read_u8_dataset(&group, "record_hashes")?;
+    if flat.len() % 32 != 0 {
+        return Err(MemoryError::Schema(
+            "/integrity/record_hashes is not a whole number of hashes".into(),
+        ));
+    }
+    let record_hashes = flat.as_chunks::<32>().0.to_vec();
+    Ok(Some(StoredSignature {
+        manifest: Manifest {
+            record_count,
+            records_root: hex::<32>(&attrs, "sig_records_root")?,
+            settings: hex::<32>(&attrs, "sig_settings")?,
+            sessions: hex::<32>(&attrs, "sig_sessions")?,
+            graph: hex::<32>(&attrs, "sig_graph")?,
+        },
+        record_hashes,
+        public_key: hex::<32>(&attrs, "sig_public_key")?,
+        signature: hex::<64>(&attrs, "sig_signature")?,
+    }))
+}
+
 /// Read the checkpoint bookkeeping from `/meta`.
 pub fn read_checkpoint_meta(file: &clawhdf5::File) -> CheckpointMeta {
-    let ann_generation = file
-        .group("meta")
+    let ann_generation =
+        meta_attrs(file)
             .ok()
-        .and_then(|g| g.attrs().ok())
             .and_then(|attrs| match attrs.get(ANN_GENERATION_ATTR) {
                 Some(AttrValue::I64(v)) => Some(*v as u64),
                 _ => None,
             });
+    let signed = meta_attrs(file).is_ok_and(|attrs| attrs.contains_key(SIG_VERSION_ATTR));
     CheckpointMeta {
         wal_applied: read_wal_mark(file),
         ann_generation,
+        signed,
     }
 }
 
@@ -440,12 +592,7 @@ pub fn validate_and_load(
     file: &clawhdf5::File,
 ) -> Result<(MemoryConfig, MemoryCache, SessionCache, KnowledgeCache), MemoryError> {
     // Read /meta group attributes
-    let meta = file
-        .group("meta")
-        .map_err(|e| MemoryError::Schema(format!("missing /meta group: {e}")))?;
-    let attrs = meta
-        .attrs()
-        .map_err(|e| MemoryError::Schema(format!("cannot read /meta attrs: {e}")))?;
+    let attrs = meta_attrs(file)?;
 
     let schema_version = match attrs.get("schema_version") {
         Some(AttrValue::String(s)) => s.clone(),
@@ -484,10 +631,32 @@ pub fn validate_and_load(
         wal_max_entries: optional_i64_attr(&attrs, "wal_max_entries")
             .and_then(|v| usize::try_from(v).ok())
             .unwrap_or(500),
+        // `false`, not the new-store default: a store written before this
+        // setting existed was built with an f32 index, and reopening it must
+        // not silently change that.
+        quantized_index: optional_bool_attr(&attrs, "quantized_index", false),
+        hnsw_m: optional_i64_attr(&attrs, "hnsw_m")
+            .and_then(|v| usize::try_from(v).ok())
+            .unwrap_or(16),
+        hnsw_ef_construction: optional_i64_attr(&attrs, "hnsw_ef_construction")
+            .and_then(|v| usize::try_from(v).ok())
+            .unwrap_or(64),
+        hnsw_ef_search: optional_i64_attr(&attrs, "hnsw_ef_search")
+            .and_then(|v| usize::try_from(v).ok())
+            .unwrap_or(0),
     };
 
     // Load /memory group
-    let memory_cache = load_memory_group(file, embedding_dim)?;
+    let mut memory_cache = load_memory_group(file, embedding_dim)?;
+    // A float16 store's cache holds half-rounded embeddings. Embeddings read
+    // from an f16 dataset already are; a float16 store whose last checkpoint
+    // predates half-precision storage is still f32 on disk and is rounded
+    // here.
+    if config.float16 && embeddings_are_f16(file) {
+        memory_cache.half_precision = true;
+    } else {
+        memory_cache.set_half_precision(config.float16);
+    }
 
     // Load /sessions group
     let session_cache = load_sessions_group(file)?;
@@ -563,12 +732,7 @@ fn load_memory_group(
             .collect(),
     };
 
-    // Unflatten embeddings
-    let embeddings: Vec<Vec<f32>> = flat_embeddings
-        .chunks(embedding_dim)
-        .map(|c| c.to_vec())
-        .collect();
-
+    // No unflattening: the cache stores the buffer as it is on disk.
     // Read activation_weights if present, default to vec![1.0; N] for backward compat
     let activation_weights = match read_f32_dataset(&group, "activation_weights") {
         Ok(w) if w.len() == n => w,
@@ -576,7 +740,7 @@ fn load_memory_group(
     };
 
     cache.chunks = chunks;
-    cache.embeddings = embeddings;
+    cache.embeddings.set_flat(embedding_dim, flat_embeddings);
     cache.source_channels = source_channels;
     cache.timestamps = timestamps;
     cache.session_ids = session_ids;
@@ -584,7 +748,6 @@ fn load_memory_group(
     cache.tombstones = tombstones;
     cache.norms = norms;
     cache.activation_weights = activation_weights;
-    cache.rebuild_flat();
 
     Ok(cache)
 }
@@ -736,6 +899,13 @@ fn read_string_dataset_from_group(
         .map_err(|e| MemoryError::Hdf5(format!("cannot read strings from {name}: {e}")))
 }
 
+/// Whether `/memory/embeddings` is stored as IEEE half precision.
+fn embeddings_are_f16(file: &clawhdf5::File) -> bool {
+    file.dataset("memory/embeddings")
+        .and_then(|ds| ds.dtype())
+        .is_ok_and(|dt| matches!(dt, clawhdf5::DType::Other(ref s) if s == "float16"))
+}
+
 fn read_f32_dataset(group: &clawhdf5::Group<'_>, name: &str) -> Result<Vec<f32>, MemoryError> {
     let ds = group
         .dataset(name)

crates/clawhdf5-agent/src/search.rs

@@ -2,18 +2,107 @@
 
 use std::path::Path;
 
+use std::collections::HashSet;
+
 use crate::bm25;
+use crate::confidence::{ConfidenceConfig, ScoredResult, reject_low_confidence};
 use crate::hybrid;
+use crate::reranker::{ReRankConfig, RerankInput, rerank};
 use crate::{HDF5Memory, MAX_ACTIVATION_WEIGHT, MemoryError, Result, SearchResult};
 
+/// Options for [`HDF5Memory::search`].
+///
+/// [`SearchOptions::new`] is plain hybrid search with the tuned default
+/// fusion — the same as `hybrid_search_with(.., hybrid::DEFAULT_FUSION, k)`.
+/// Every stage beyond that is opt-in.
+#[derive(Debug, Clone)]
+pub struct SearchOptions {
+    /// Number of results to return.
+    pub k: usize,
+    /// How the vector and keyword stages are combined.
+    pub fusion: hybrid::Fusion,
+    /// Only consider records whose `source_channel` is one of these. The
+    /// filter applies *before* ranking, so a filtered search still returns up
+    /// to `k` results and scores are normalised over the records it can
+    /// return. `None` searches everything; an empty list matches nothing.
+    pub source_channels: Option<Vec<String>>,
+    /// Re-rank a candidate pool by retrieval relevance, recency, source
+    /// authority and activation — the pipeline the OpenClaw backend runs.
+    pub rerank: Option<ReRankConfig>,
+    /// Candidates retrieved for re-ranking; 0 means `max(3k, 10)`.
+    pub rerank_pool: usize,
+    /// Drop low-confidence results (after re-ranking, when that is on).
+    pub confidence: Option<ConfidenceConfig>,
+    /// The time recency is measured from, in seconds since the epoch.
+    /// `None` uses the system clock.
+    pub now: Option<f64>,
+}
+
+impl SearchOptions {
+    pub fn new(k: usize) -> Self {
+        Self {
+            k,
+            fusion: hybrid::DEFAULT_FUSION,
+            source_channels: None,
+            rerank: None,
+            rerank_pool: 0,
+            confidence: None,
+            now: None,
+        }
+    }
+
+    pub fn with_fusion(mut self, fusion: hybrid::Fusion) -> Self {
+        self.fusion = fusion;
+        self
+    }
+
+    /// Search only records from these source channels.
+    pub fn with_sources<S: Into<String>>(mut self, channels: impl IntoIterator<Item = S>) -> Self {
+        self.source_channels = Some(channels.into_iter().map(Into::into).collect());
+        self
+    }
+
+    pub fn with_rerank(mut self, config: ReRankConfig) -> Self {
+        self.rerank = Some(config);
+        self
+    }
+
+    pub fn with_confidence(mut self, config: ConfidenceConfig) -> Self {
+        self.confidence = Some(config);
+        self
+    }
+
+    /// Measure recency from `now` (seconds since the epoch) instead of the
+    /// system clock — for reproducible results and tests.
+    pub fn at_time(mut self, now: f64) -> Self {
+        self.now = Some(now);
+        self
+    }
+}
+
+impl Default for SearchOptions {
+    fn default() -> Self {
+        Self::new(10)
+    }
+}
+
 impl HDF5Memory {
-    /// Vector + keyword scoring stage of [`HDF5Memory::hybrid_search`].
+    /// Vector + keyword scoring stage of [`HDF5Memory::search`].
     ///
     /// Without the `hnsw` feature this is a full linear cosine scan (the exact
     /// previous behaviour, also used as the correctness oracle in tests). With
     /// `hnsw` enabled and an index available, the vector candidates come from an
     /// approximate-nearest-neighbour search over an over-fetched pool, then merge
     /// with BM25 via the shared [`hybrid::merge_vector_keyword`].
+    ///
+    /// `exclude`, when given, marks records that must not be returned (1 =
+    /// excluded; it covers tombstones too). The index is over-fetched in
+    /// proportion to how much the mask removes. Surfacing `pool` candidates
+    /// costs the index roughly `pool × M` distance evaluations, while an exact
+    /// scan of the allowed records costs one each — so whenever that scan is
+    /// the cheaper of the two it is used instead, and it is also the fallback
+    /// if the pool comes back with too few allowed hits (the allowed records
+    /// sit away from the query). A filtered search never comes back short.
     #[cfg(feature = "hnsw")]
     fn vector_keyword_search(
         &mut self,
@@ -22,24 +111,103 @@ impl HDF5Memory {
         bm25: &bm25::BM25Index,
         fusion: hybrid::Fusion,
         k: usize,
+        exclude: Option<&[u8]>,
     ) -> Vec<(usize, f32)> {
         self.ensure_hnsw_fresh();
+        let n = self.cache.len();
+        // Over-fetch so the merge sees a useful vector pool. `ef` is
+        // configurable, but the pool the fusion stage sees is not tied to it:
+        // a caller lowering `ef` for speed should not silently narrow what
+        // fusion has to work with.
+        let mut pool = (k * 8).max(64);
+        let mut allowed = n;
+        if let Some(ex) = exclude {
+            allowed = ex.iter().filter(|&&e| e == 0).count();
+            if allowed == 0 {
+                return Vec::new();
+            }
+            // Expect `pool` allowed hits if the filter is independent of the
+            // query's neighbourhood.
+            pool = pool.saturating_mul(n).div_ceil(allowed);
+            if allowed <= pool.saturating_mul(self.hnsw_m()) {
+                return self.exact_masked_search(query_embedding, query_text, bm25, fusion, k, ex);
+            }
+        }
         match self.hnsw.as_ref() {
             Some(index) if !index.is_empty() && index.dimension() == query_embedding.len() => {
-                // Over-fetch so the merge sees a useful vector pool; cosine
-                // distance from the index converts back to similarity (1 - d).
-                let pool = (k * 8).max(64);
-                let vec_scores: Vec<(usize, f32)> = index
-                    .search(query_embedding, pool, pool)
+                let ef = self.hnsw_ef_search(k).max(pool);
+                let candidates = index.search(query_embedding, pool, ef);
+                // A quantised index returns approximate distances, and no
+                // amount of `ef` fixes that — the loss is in the distances,
+                // not the graph. Re-score the pool against the cache's exact
+                // embeddings, which cost nothing extra to keep: recall then
+                // matches an f32 index. See `BENCHMARKS.md`.
+                let exact = index.storage() == clawhdf5_ann::Storage::Int8;
+                let vec_scores: Vec<(usize, f32)> = candidates
                     .into_iter()
-                    .map(|(id, dist)| (id, 1.0 - dist))
+                    .filter(|(id, _)| exclude.is_none_or(|ex| ex[*id] == 0))
+                    .map(|(id, dist)| {
+                        let score = if exact {
+                            crate::vector_search::cosine_similarity(
+                                query_embedding,
+                                &self.cache.embeddings[id],
+                            )
+                        } else {
+                            1.0 - dist
+                        };
+                        (id, score)
+                    })
                     .collect();
                 // Fusion normalises over every keyword match, so it needs all
                 // the scores — but not ranked.
-                let kw_scores = bm25.scores(query_text);
+                let mut kw_scores = bm25.scores(query_text);
+                if let Some(ex) = exclude {
+                    if vec_scores.len() < k.min(allowed) {
+                        // The allowed records are not where the index looked.
+                        return self.exact_masked_search(
+                            query_embedding,
+                            query_text,
+                            bm25,
+                            fusion,
+                            k,
+                            ex,
+                        );
+                    }
+                    kw_scores.retain(|(id, _)| ex[*id] == 0);
+                }
                 hybrid::fuse(vec_scores, kw_scores, fusion, k)
             }
-            _ => hybrid::hybrid_search_fused(
+            _ => match exclude {
+                Some(ex) => {
+                    self.exact_masked_search(query_embedding, query_text, bm25, fusion, k, ex)
+                }
+                None => hybrid::hybrid_search_fused(
+                    query_embedding,
+                    query_text,
+                    &self.cache.embeddings,
+                    &self.cache.chunks,
+                    &self.cache.tombstones,
+                    bm25,
+                    fusion,
+                    k,
+                ),
+            },
+        }
+    }
+
+    #[cfg(not(feature = "hnsw"))]
+    fn vector_keyword_search(
+        &mut self,
+        query_embedding: &[f32],
+        query_text: &str,
+        bm25: &bm25::BM25Index,
+        fusion: hybrid::Fusion,
+        k: usize,
+        exclude: Option<&[u8]>,
+    ) -> Vec<(usize, f32)> {
+        match exclude {
+            Some(ex) => self.exact_masked_search(query_embedding, query_text, bm25, fusion, k, ex),
+            None => hybrid::hybrid_search_fused(
                 query_embedding,
                 query_text,
                 &self.cache.embeddings,
@@ -52,25 +220,33 @@ impl HDF5Memory {
         }
     }
 
-    #[cfg(not(feature = "hnsw"))]
-    fn vector_keyword_search(
-        &mut self,
+    /// Exact hybrid search over the records `exclude` leaves (0 = allowed).
+    fn exact_masked_search(
+        &self,
         query_embedding: &[f32],
         query_text: &str,
         bm25: &bm25::BM25Index,
         fusion: hybrid::Fusion,
         k: usize,
+        exclude: &[u8],
     ) -> Vec<(usize, f32)> {
-        hybrid::hybrid_search_fused(
-            query_embedding,
-            query_text,
-            &self.cache.embeddings,
-            &self.cache.chunks,
-            &self.cache.tombstones,
-            bm25,
-            fusion,
-            k,
-        )
+        let vec_scores =
+            hybrid::exact_vector_scores(query_embedding, &self.cache.embeddings, exclude);
+        let mut kw_scores = bm25.scores(query_text);
+        kw_scores.retain(|(id, _)| exclude.get(*id) == Some(&0));
+        hybrid::fuse(vec_scores, kw_scores, fusion, k)
+    }
+
+    /// The exclusion mask for a source-channel filter: 1 for a tombstoned
+    /// record or one from a channel not in `channels`.
+    fn source_mask(&self, channels: &[String]) -> Vec<u8> {
+        let allowed: HashSet<&str> = channels.iter().map(String::as_str).collect();
+        self.cache
+            .source_channels
+            .iter()
+            .zip(&self.cache.tombstones)
+            .map(|(ch, &t)| u8::from(t != 0 || !allowed.contains(ch.as_str())))
+            .collect()
     }
 
     /// Perform hybrid search combining cosine vector similarity and BM25 keyword search.
@@ -105,12 +281,53 @@ impl HDF5Memory {
         fusion: hybrid::Fusion,
         k: usize,
     ) -> Vec<SearchResult> {
+        self.search(
+            query_embedding,
+            query_text,
+            &SearchOptions::new(k).with_fusion(fusion),
+        )
+    }
+
+    /// Hybrid search with optional source filtering, re-ranking and
+    /// confidence rejection — see [`SearchOptions`].
+    ///
+    /// Stages, in order: vector + keyword retrieval over the records the
+    /// source filter allows; fusion; scaling by Hebbian activation; re-ranking
+    /// (if on) of a `rerank_pool` of candidates; confidence rejection (if on);
+    /// the top `k`. The records returned with a positive score get their
+    /// Hebbian boost.
+    pub fn search(
+        &mut self,
+        query_embedding: &[f32],
+        query_text: &str,
+        options: &SearchOptions,
+    ) -> Vec<SearchResult> {
+        let k = options.k;
+        let fetch = match options.rerank {
+            Some(_) if options.rerank_pool > 0 => options.rerank_pool.max(k),
+            Some(_) => k.saturating_mul(3).max(10),
+            None => k,
+        };
+        let exclude = options
+            .source_channels
+            .as_deref()
+            .map(|channels| self.source_mask(channels));
+
         // The keyword index lives for the life of the store and is updated
         // incrementally. Take it out for the duration of the call so the
         // vector stage can borrow `self` mutably, then put it back.
         self.ensure_bm25_fresh();
         let bm25 = self.bm25.take().expect("ensure_bm25_fresh leaves an index");
-        let scored = self.vector_keyword_search(query_embedding, query_text, &bm25, fusion, k);
+        let scored = self.vector_keyword_search(
+            query_embedding,
+            query_text,
+            &bm25,
+            options.fusion,
+            fetch,
+            exclude.as_deref(),
+        );
+        self.bm25 = Some(bm25);
+
         let mut results: Vec<SearchResult> = scored
             .into_iter()
             .map(|(idx, score)| {
@@ -134,6 +351,25 @@ impl HDF5Memory {
                 .then(a.index.cmp(&b.index))
         });
 
+        if let Some(config) = &options.rerank {
+            results = Self::rerank_results(results, config, options.now);
+        }
+        if let Some(config) = &options.confidence {
+            let scored: Vec<ScoredResult> = results
+                .iter()
+                .map(|r| ScoredResult {
+                    index: r.index,
+                    score: r.score,
+                })
+                .collect();
+            let keep: HashSet<usize> = reject_low_confidence(&scored, config)
+                .into_iter()
+                .map(|r| r.index)
+                .collect();
+            results.retain(|r| keep.contains(&r.index));
+        }
+        results.truncate(k);
+
         // Only reinforce records that actually matched. When fewer than `k`
         // records are relevant, the rest of the list is zero-score filler;
         // boosting it would teach the store that arbitrary records are
@@ -144,11 +380,45 @@ impl HDF5Memory {
             .map(|r| r.index)
             .collect();
         self.apply_hebbian_boost(&hit_indices);
-        self.bm25 = Some(bm25);
 
         results
     }
 
+    /// Reorder by the re-ranker's combined score, which also becomes each
+    /// result's `score`.
+    fn rerank_results(
+        results: Vec<SearchResult>,
+        config: &ReRankConfig,
+        now: Option<f64>,
+    ) -> Vec<SearchResult> {
+        let now = now.unwrap_or_else(|| {
+            std::time::SystemTime::now()
+                .duration_since(std::time::UNIX_EPOCH)
+                .map(|d| d.as_secs_f64())
+                .unwrap_or(0.0)
+        });
+        let inputs: Vec<RerankInput> = results
+            .iter()
+            .map(|r| RerankInput {
+                index: r.index,
+                timestamp: r.timestamp,
+                source_channel: r.source_channel.clone(),
+                raw_activation: r.activation,
+                relevance: r.score,
+            })
+            .collect();
+        let mut by_index: std::collections::HashMap<usize, SearchResult> =
+            results.into_iter().map(|r| (r.index, r)).collect();
+        rerank(&inputs, config, now)
+            .into_iter()
+            .filter_map(|rr| {
+                let mut r = by_index.remove(&rr.index)?;
+                r.score = rr.combined_score;
+                Some(r)
+            })
+            .collect()
+    }
+
     /// Reinforce the records a query returned. The new weights are persisted by
     /// the next checkpoint (any write that flushes, `flush_wal`, or drop) — not
     /// by rewriting the whole store inside the query, which is what made

crates/clawhdf5-agent/src/session.rs

@@ -33,7 +33,7 @@ impl SessionCache {
         self.entries.is_empty()
     }
 
-    /// Add a new session with its summary.
+    /// Add a new session with its summary, timestamped now.
     pub fn add(
         &mut self,
         id: &str,
@@ -47,6 +47,21 @@ impl SessionCache {
             .unwrap_or_default()
             .as_secs_f64()
             * 1_000_000.0; // microseconds
+        self.add_at(id, start_idx, end_idx, channel, summary, ts);
+    }
+
+    /// Add a session with an explicit timestamp (Unix **microseconds**, the
+    /// unit [`SessionEntry::ts`] uses) — for importers carrying sessions over
+    /// from another store, whose original time should be kept.
+    pub fn add_at(
+        &mut self,
+        id: &str,
+        start_idx: usize,
+        end_idx: usize,
+        channel: &str,
+        summary: &str,
+        ts: f64,
+    ) {
         self.entries.push(SessionEntry {
             id: id.to_string(),
             start_idx: start_idx as u64,

crates/clawhdf5-agent/src/signing.rs

@@ -0,0 +1,419 @@
+//! Ed25519-signed checkpoints.
+//!
+//! When a signing key is set ([`crate::HDF5Memory::set_signing_key`]), every
+//! checkpoint writes a signed manifest of the store: a SHA-256 per memory
+//! record rolled into a Merkle root, plus hashes of the store's settings, its
+//! sessions and its knowledge graph. [`verify_store`] recomputes all of it from
+//! the file and checks the signature against a public key the caller trusts,
+//! so any change to the checkpointed file — a record's text or embedding, a
+//! setting, a session, a graph edge, made through this crate or any other HDF5
+//! tool — is detected, and the per-record hashes say which records changed.
+//!
+//! What it does not cover: saves still only in the WAL (made since the last
+//! checkpoint). [`VerifyReport::wal_entries_unsigned`] counts them.
+//!
+//! The hashes cover exactly what the file persists, in the form the loader
+//! returns it, so a store verifies after any number of reopen/checkpoint
+//! cycles. Derived data (L2 norms, the vector index) is not covered; it is
+//! recomputed from covered data.
+
+use ed25519_dalek::{Signature, Signer, Verifier};
+pub use ed25519_dalek::{SigningKey, VerifyingKey};
+use sha2::{Digest, Sha256};
+
+use crate::MemoryConfig;
+use crate::cache::MemoryCache;
+use crate::knowledge::KnowledgeCache;
+use crate::session::SessionCache;
+use crate::wal::WalMark;
+
+/// Version of the manifest encoding; part of what is signed.
+pub const MANIFEST_VERSION: i64 = 1;
+
+type Hash = [u8; 32];
+
+/// The hashes a signature covers.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct Manifest {
+    pub record_count: u64,
+    /// Merkle root over the per-record hashes.
+    pub records_root: Hash,
+    /// Settings persisted in `/meta`, plus the checkpoint's WAL mark.
+    pub settings: Hash,
+    pub sessions: Hash,
+    pub graph: Hash,
+}
+
+impl Manifest {
+    /// The exact bytes that are signed.
+    pub fn signed_bytes(&self) -> Vec<u8> {
+        let mut m = Vec::with_capacity(160);
+        m.extend_from_slice(b"clawhdf5-agent signed checkpoint\0");
+        m.extend_from_slice(&MANIFEST_VERSION.to_le_bytes());
+        m.extend_from_slice(&self.record_count.to_le_bytes());
+        m.extend_from_slice(&self.records_root);
+        m.extend_from_slice(&self.settings);
+        m.extend_from_slice(&self.sessions);
+        m.extend_from_slice(&self.graph);
+        m
+    }
+}
+
+/// A signature as stored in a checkpoint.
+#[derive(Debug, Clone)]
+pub struct StoredSignature {
+    pub manifest: Manifest,
+    pub record_hashes: Vec<Hash>,
+    pub public_key: [u8; 32],
+    pub signature: [u8; 64],
+}
+
+/// Build the manifest (and per-record hashes) for the state about to be
+/// checkpointed, and sign it.
+pub fn sign(
+    key: &SigningKey,
+    config: &MemoryConfig,
+    cache: &MemoryCache,
+    sessions: &SessionCache,
+    knowledge: &KnowledgeCache,
+    wal_applied: Option<WalMark>,
+) -> StoredSignature {
+    let (manifest, record_hashes) = manifest(config, cache, sessions, knowledge, wal_applied);
+    let signature = key.sign(&manifest.signed_bytes()).to_bytes();
+    StoredSignature {
+        manifest,
+        record_hashes,
+        public_key: key.verifying_key().to_bytes(),
+        signature,
+    }
+}
+
+/// Compute the manifest of a store's state.
+pub fn manifest(
+    config: &MemoryConfig,
+    cache: &MemoryCache,
+    sessions: &SessionCache,
+    knowledge: &KnowledgeCache,
+    wal_applied: Option<WalMark>,
+) -> (Manifest, Vec<Hash>) {
+    let record_hashes: Vec<Hash> = (0..cache.len()).map(|i| record_hash(cache, i)).collect();
+    let manifest = Manifest {
+        record_count: cache.len() as u64,
+        records_root: merkle_root(&record_hashes),
+        settings: settings_hash(config, wal_applied),
+        sessions: sessions_hash(sessions),
+        graph: graph_hash(knowledge),
+    };
+    (manifest, record_hashes)
+}
+
+// ---------------------------------------------------------------------------
+// Canonical encoding
+// ---------------------------------------------------------------------------
+
+/// A SHA-256 over length-prefixed fields, so no two different field lists
+/// hash the same bytes.
+struct Fields(Sha256);
+
+impl Fields {
+    fn new(domain: &str) -> Self {
+        let mut h = Sha256::new();
+        h.update((domain.len() as u64).to_le_bytes());
+        h.update(domain.as_bytes());
+        Self(h)
+    }
+    fn bytes(&mut self, b: &[u8]) -> &mut Self {
+        self.0.update((b.len() as u64).to_le_bytes());
+        self.0.update(b);
+        self
+    }
+    /// Strings as the loader returns them: stored null-padded, so a trailing
+    /// NUL cannot survive a round trip and must not be part of the hash.
+    fn str(&mut self, s: &str) -> &mut Self {
+        self.bytes(s.trim_end_matches('\0').as_bytes())
+    }
+    fn u64(&mut self, v: u64) -> &mut Self {
+        self.0.update(v.to_le_bytes());
+        self
+    }
+    fn f64(&mut self, v: f64) -> &mut Self {
+        self.0.update(v.to_bits().to_le_bytes());
+        self
+    }
+    fn f32(&mut self, v: f32) -> &mut Self {
+        self.0.update(v.to_bits().to_le_bytes());
+        self
+    }
+    fn finish(self) -> Hash {
+        self.0.finalize().into()
+    }
+}
+
+/// Everything persisted about record `i`, including its position. The
+/// embedding is hashed as the cache holds it — for a `float16` store that is
+/// the half-rounded value the file holds.
+fn record_hash(cache: &MemoryCache, i: usize) -> Hash {
+    let mut f = Fields::new("clawhdf5-agent/record");
+    f.u64(i as u64).str(&cache.chunks[i]);
+    let emb: Vec<u8> = cache.embeddings[i]
+        .iter()
+        .flat_map(|v| v.to_bits().to_le_bytes())
+        .collect();
+    f.bytes(&emb)
+        .str(&cache.source_channels[i])
+        .f64(cache.timestamps[i])
+        .str(&cache.session_ids[i])
+        .str(&cache.tags[i])
+        .u64(u64::from(cache.tombstones[i]))
+        .f32(cache.activation_weights[i]);
+    f.finish()
+}
+
+/// Binary Merkle tree: leaves are the record hashes; a parent hashes its two
+/// children with a node prefix; an odd node is carried up unchanged.
+fn merkle_root(leaves: &[Hash]) -> Hash {
+    if leaves.is_empty() {
+        return Fields::new("clawhdf5-agent/merkle-empty").finish();
+    }
+    let mut level: Vec<Hash> = leaves.to_vec();
+    while level.len() > 1 {
+        level = level
+            .chunks(2)
+            .map(|pair| match pair {
+                [l, r] => {
+                    let mut h = Sha256::new();
+                    h.update([1u8]);
+                    h.update(l);
+                    h.update(r);
+                    h.finalize().into()
+                }
+                [only] => *only,
+                _ => unreachable!(),
+            })
+            .collect();
+    }
+    level[0]
+}
+
+fn settings_hash(c: &MemoryConfig, wal_applied: Option<WalMark>) -> Hash {
+    let mut f = Fields::new("clawhdf5-agent/settings");
+    f.str(crate::schema::SCHEMA_VERSION)
+        .str(&c.created_at)
+        .str(&c.agent_id)
+        .str(&c.embedder)
+        .u64(c.embedding_dim as u64)
+        .u64(c.chunk_size as u64)
+        .u64(c.overlap as u64)
+        .u64(u64::from(c.float16))
+        .u64(u64::from(c.compression))
+        .u64(u64::from(c.compression_level))
+        .f32(c.compact_threshold)
+        .f32(c.hebbian_boost)
+        .f32(c.decay_factor)
+        .u64(u64::from(c.wal_enabled))
+        .u64(c.wal_max_entries as u64)
+        .u64(u64::from(c.quantized_index))
+        .u64(c.hnsw_m as u64)
+        .u64(c.hnsw_ef_construction as u64)
+        .u64(c.hnsw_ef_search as u64);
+    // An empty mark is not written to the file, so it must hash as none.
+    match wal_applied.filter(|m| m.len > 0) {
+        Some(m) => f.u64(1).u64(m.len).u64(u64::from(m.crc)),
+        None => f.u64(0),
+    };
+    f.finish()
+}
+
+fn sessions_hash(s: &SessionCache) -> Hash {
+    let mut f = Fields::new("clawhdf5-agent/sessions");
+    f.u64(s.entries.len() as u64);
+    for (i, e) in s.entries.iter().enumerate() {
+        f.str(&e.id)
+            .u64(e.start_idx)
+            .u64(e.end_idx)
+            .str(&e.channel)
+            .f64(e.ts)
+            .str(s.summaries.get(i).map(String::as_str).unwrap_or(""));
+    }
+    f.finish()
+}
+
+fn graph_hash(k: &KnowledgeCache) -> Hash {
+    let mut f = Fields::new("clawhdf5-agent/graph");
+    f.u64(k.entities.len() as u64);
+    for e in &k.entities {
+        f.u64(e.id)
+            .str(&e.name)
+            .str(&e.entity_type)
+            .u64(e.embedding_idx as u64);
+    }
+    f.u64(k.relations.len() as u64);
+    for r in &k.relations {
+        f.u64(r.src)
+            .u64(r.tgt)
+            .str(&r.relation)
+            .f32(r.weight)
+            .f64(r.ts);
+    }
+    f.u64(k.alias_strings.len() as u64);
+    for (s, id) in k.alias_strings.iter().zip(&k.alias_entity_ids) {
+        f.str(s).u64(*id as u64);
+    }
+    f.finish()
+}
+
+// ---------------------------------------------------------------------------
+// Verification
+// ---------------------------------------------------------------------------
+
+/// The outcome of [`verify_store`].
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct VerifyReport {
+    /// The checkpoint carries a signature.
+    pub signed: bool,
+    /// The signature was made by the key the caller trusts.
+    pub key_matches: bool,
+    /// The signature over the stored manifest is valid.
+    pub signature_valid: bool,
+    /// The file's current contents match the signed manifest.
+    pub records_match: bool,
+    pub settings_match: bool,
+    pub sessions_match: bool,
+    pub graph_match: bool,
+    /// Records whose contents differ from what was signed (by position),
+    /// when the stored per-record hashes are themselves authentic.
+    pub changed_records: Vec<usize>,
+    /// Records in the file versus in the signed manifest.
+    pub record_count: u64,
+    pub signed_record_count: u64,
+    /// The public key the checkpoint claims to be signed by.
+    pub public_key: Option<[u8; 32]>,
+    /// Saves in the WAL after the checkpoint: not covered by the signature.
+    pub wal_entries_unsigned: usize,
+}
+
+impl VerifyReport {
+    /// Signed by the trusted key, signature valid, and every part of the
+    /// file unchanged since it was signed.
+    pub fn is_valid(&self) -> bool {
+        self.signed
+            && self.key_matches
+            && self.signature_valid
+            && self.records_match
+            && self.settings_match
+            && self.sessions_match
+            && self.graph_match
+    }
+}
+
+/// Check a store file against the public key the caller trusts.
+///
+/// Reads the checkpoint (not the WAL), recomputes every hash from its
+/// contents and checks the signature. Never writes.
+pub fn verify_store(
+    path: &std::path::Path,
+    trusted: &VerifyingKey,
+) -> Result<VerifyReport, crate::MemoryError> {
+    let file = clawhdf5::File::open(path)
+        .map_err(|e| crate::MemoryError::Hdf5(format!("cannot open {}: {e}", path.display())))?;
+    let (config, cache, sessions, knowledge) = crate::schema::validate_and_load(&file)?;
+    let checkpoint = crate::schema::read_checkpoint_meta(&file);
+    let stored = crate::schema::read_signature(&file)?;
+    let wal_entries_unsigned = count_wal_entries_after(path, checkpoint.wal_applied);
+
+    let (current, current_hashes) = manifest(
+        &config,
+        &cache,
+        &sessions,
+        &knowledge,
+        checkpoint.wal_applied,
+    );
+
+    let Some(stored) = stored else {
+        return Ok(VerifyReport {
+            signed: false,
+            key_matches: false,
+            signature_valid: false,
+            records_match: false,
+            settings_match: false,
+            sessions_match: false,
+            graph_match: false,
+            changed_records: Vec::new(),
+            record_count: current.record_count,
+            signed_record_count: 0,
+            public_key: None,
+            wal_entries_unsigned,
+        });
+    };
+
+    let key_matches = stored.public_key == trusted.to_bytes();
+    let signature_valid = trusted
+        .verify(
+            &stored.manifest.signed_bytes(),
+            &Signature::from_bytes(&stored.signature),
+        )
+        .is_ok();
+    // The stored per-record hashes can localise a change only if they are
+    // the ones that were signed.
+    let hashes_authentic = signature_valid
+        && stored.record_hashes.len() as u64 == stored.manifest.record_count
+        && merkle_root(&stored.record_hashes) == stored.manifest.records_root;
+    let changed_records = if hashes_authentic {
+        let n = current_hashes.len().max(stored.record_hashes.len());
+        (0..n)
+            .filter(|&i| current_hashes.get(i) != stored.record_hashes.get(i))
+            .collect()
+    } else {
+        Vec::new()
+    };
+
+    Ok(VerifyReport {
+        signed: true,
+        key_matches,
+        signature_valid,
+        records_match: signature_valid
+            && current.record_count == stored.manifest.record_count
+            && current.records_root == stored.manifest.records_root,
+        settings_match: signature_valid && current.settings == stored.manifest.settings,
+        sessions_match: signature_valid && current.sessions == stored.manifest.sessions,
+        graph_match: signature_valid && current.graph == stored.manifest.graph,
+        changed_records,
+        record_count: current.record_count,
+        signed_record_count: stored.manifest.record_count,
+        public_key: Some(stored.public_key),
+        wal_entries_unsigned,
+    })
+}
+
+fn count_wal_entries_after(store: &std::path::Path, mark: Option<WalMark>) -> usize {
+    let wal = store.with_extension("h5.wal");
+    if !wal.exists() {
+        return 0;
+    }
+    crate::wal::WalFile::read_entries_for_migration(&wal, mark)
+        .map(|e| e.len())
+        .unwrap_or(0)
+}
+
+/// A new random signing key from the operating system's RNG.
+pub fn generate_key() -> SigningKey {
+    SigningKey::generate(&mut rand_core::OsRng)
+}
+
+/// Hex encoding for keys and signatures in attributes and the CLI.
+pub fn to_hex(bytes: &[u8]) -> String {
+    bytes.iter().map(|b| format!("{b:02x}")).collect()
+}
+
+/// Parse hex into exactly `N` bytes.
+pub fn from_hex<const N: usize>(s: &str) -> Option<[u8; N]> {
+    let s = s.trim();
+    if s.len() != 2 * N {
+        return None;
+    }
+    let mut out = [0u8; N];
+    for (i, byte) in out.iter_mut().enumerate() {
+        *byte = u8::from_str_radix(&s[2 * i..2 * i + 2], 16).ok()?;
+    }
+    Some(out)
+}

crates/clawhdf5-agent/src/storage.rs

@@ -36,7 +36,7 @@ pub fn write_to_disk_with_mark(
 ) -> Result<(), MemoryError> {
     let meta = schema::CheckpointMeta {
         wal_applied,
-        ann_generation: None,
+        ..schema::CheckpointMeta::default()
     };
     write_to_disk_with_meta(path, config, cache, sessions, knowledge, &meta)
 }
@@ -50,7 +50,21 @@ pub fn write_to_disk_with_meta(
     knowledge: &KnowledgeCache,
     checkpoint: &schema::CheckpointMeta,
 ) -> Result<(), MemoryError> {
-    let bytes = schema::build_hdf5_file_with_meta(config, cache, sessions, knowledge, checkpoint)?;
+    write_to_disk_signed(path, config, cache, sessions, knowledge, checkpoint, None)
+}
+
+/// [`write_to_disk_with_meta`] with a signed manifest of the contents.
+pub fn write_to_disk_signed(
+    path: &Path,
+    config: &MemoryConfig,
+    cache: &MemoryCache,
+    sessions: &SessionCache,
+    knowledge: &KnowledgeCache,
+    checkpoint: &schema::CheckpointMeta,
+    signature: Option<&crate::signing::StoredSignature>,
+) -> Result<(), MemoryError> {
+    let bytes =
+        schema::build_hdf5_file_signed(config, cache, sessions, knowledge, checkpoint, signature)?;
 
     if bytes.is_empty() {
         return Err(MemoryError::Hdf5("build_hdf5_file produced 0 bytes".into()));
@@ -113,13 +127,11 @@ pub type StoreState = (MemoryConfig, MemoryCache, SessionCache, KnowledgeCache);
 /// [`read_from_disk`], plus the checkpoint's [`WalMark`] (if any) so the
 /// caller can skip WAL entries this file already contains.
 pub fn read_from_disk_with_mark(path: &Path) -> Result<(StoreState, Option<WalMark>), MemoryError> {
-    let mmap = clawhdf5_io::MmapReader::open(path).map_err(MemoryError::Io)?;
-
-    // Advise the OS we'll need the whole file for parsing
-    mmap.advise_willneed(0, mmap.len());
-
-    // Parse the HDF5 file from the mmap'd bytes
-    let file = clawhdf5::File::from_bytes(mmap.as_bytes().to_vec())
+    // `File::open` memory-maps the file itself (the facade's `mmap` feature is
+    // on by default). Mapping it here and handing over `as_bytes().to_vec()`
+    // did the same work and then copied the whole store — a second full copy
+    // of the file, live for the whole parse, on top of the mapping.
+    let file = clawhdf5::File::open(path)
         .map_err(|e| MemoryError::Hdf5(format!("cannot open {}: {e}", path.display())))?;
 
     let (mut config, cache, sessions, knowledge) = schema::validate_and_load(&file)?;
@@ -133,9 +145,7 @@ pub fn read_from_disk_with_mark(path: &Path) -> Result<(StoreState, Option<WalMa
 pub fn read_from_disk_with_meta(
     path: &Path,
 ) -> Result<(StoreState, schema::CheckpointMeta), MemoryError> {
-    let mmap = clawhdf5_io::MmapReader::open(path).map_err(MemoryError::Io)?;
-    mmap.advise_willneed(0, mmap.len());
-    let file = clawhdf5::File::from_bytes(mmap.as_bytes().to_vec())
+    let file = clawhdf5::File::open(path)
         .map_err(|e| MemoryError::Hdf5(format!("cannot open {}: {e}", path.display())))?;
     let (mut config, cache, sessions, knowledge) = schema::validate_and_load(&file)?;
     config.path = path.to_path_buf();

crates/clawhdf5-agent/src/store_lock.rs

@@ -20,7 +20,7 @@ const LOCK_RETRY_DELAY: std::time::Duration = std::time::Duration::from_millis(1
 /// never leaves a stale lock behind; the empty lock file itself is harmless).
 #[derive(Debug)]
 pub(crate) struct StoreLock {
-    _file: File,
+    file: File,
 }
 
 impl StoreLock {
@@ -42,7 +42,7 @@ impl StoreLock {
         let mut attempts_left = LOCK_RETRIES;
         loop {
             match file.try_lock() {
-                Ok(()) => return Ok(Self { _file: file }),
+                Ok(()) => return Ok(Self { file }),
                 Err(TryLockError::WouldBlock) if attempts_left > 0 => {
                     attempts_left -= 1;
                     std::thread::sleep(LOCK_RETRY_DELAY);
@@ -60,6 +60,16 @@ impl StoreLock {
     }
 }
 
+impl Drop for StoreLock {
+    /// Unlocks before the file is closed: a process another thread forks
+    /// inherits the descriptor until it execs, and a `flock` lasts while any
+    /// descriptor of the open file does, so closing alone could keep the
+    /// store locked for a moment after the drop (see `FileEditor`'s `Drop`).
+    fn drop(&mut self) {
+        let _ = self.file.unlock();
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;

crates/clawhdf5-agent/src/vector_search.rs

@@ -4,6 +4,44 @@
 //! `clawhdf5_accel`, with optional float16 support via the `half` crate.
 //! Supports pre-computed norms for eliminating redundant norm computations.
 
+/// A corpus of equal-length embeddings addressable by index.
+///
+/// Lets the batch kernels read either the cache's flat `[N x dim]` buffer or a
+/// plain `Vec<Vec<f32>>` without either side owning a second copy.
+pub trait VectorSet {
+    /// Number of embeddings.
+    fn count(&self) -> usize;
+    /// Embedding `i`; callers only index below [`VectorSet::count`].
+    fn row(&self, i: usize) -> &[f32];
+}
+
+impl VectorSet for [Vec<f32>] {
+    fn count(&self) -> usize {
+        self.len()
+    }
+    fn row(&self, i: usize) -> &[f32] {
+        &self[i]
+    }
+}
+
+impl VectorSet for Vec<Vec<f32>> {
+    fn count(&self) -> usize {
+        self.len()
+    }
+    fn row(&self, i: usize) -> &[f32] {
+        &self[i]
+    }
+}
+
+impl VectorSet for crate::cache::Embeddings {
+    fn count(&self) -> usize {
+        self.len()
+    }
+    fn row(&self, i: usize) -> &[f32] {
+        &self[i]
+    }
+}
+
 /// Compute cosine similarity between two f32 slices.
 ///
 /// Returns 0.0 if either vector has zero magnitude.
@@ -22,7 +60,7 @@ pub fn cosine_similarity(a: &[f32], b: &[f32]) -> f32 {
 /// Returns `(index, score)` pairs sorted by score descending.
 pub fn cosine_similarity_batch(
     query: &[f32],
-    vectors: &[Vec<f32>],
+    vectors: &(impl VectorSet + ?Sized),
     tombstones: &[u8],
 ) -> Vec<(usize, f32)> {
     let query_norm = clawhdf5_accel::vector_norm(query);
@@ -30,7 +68,7 @@ pub fn cosine_similarity_batch(
         return Vec::new();
     }
 
-    let n = vectors.len();
+    let n = vectors.count();
     let mut results: Vec<(usize, f32)> = Vec::with_capacity(n);
 
     // Process 4 vectors at a time where possible
@@ -42,8 +80,9 @@ pub fn cosine_similarity_batch(
             if i < tombstones.len() && tombstones[i] != 0 {
                 continue;
             }
-            let vec_norm = clawhdf5_accel::vector_norm(&vectors[i]);
-            let score = crate::cosine_similarity_prenorm(query, query_norm, &vectors[i], vec_norm);
+            let vec_norm = clawhdf5_accel::vector_norm(vectors.row(i));
+            let score =
+                crate::cosine_similarity_prenorm(query, query_norm, vectors.row(i), vec_norm);
             results.push((i, score));
         }
     }
@@ -53,8 +92,8 @@ pub fn cosine_similarity_batch(
         if i < tombstones.len() && tombstones[i] != 0 {
             continue;
         }
-        let vec_norm = clawhdf5_accel::vector_norm(&vectors[i]);
-        let score = crate::cosine_similarity_prenorm(query, query_norm, &vectors[i], vec_norm);
+        let vec_norm = clawhdf5_accel::vector_norm(vectors.row(i));
+        let score = crate::cosine_similarity_prenorm(query, query_norm, vectors.row(i), vec_norm);
         results.push((i, score));
     }
 
@@ -68,7 +107,7 @@ pub fn cosine_similarity_batch(
 /// collections. Uses `score = dot(query, vec) / (query_norm * stored_norm)`.
 pub fn cosine_similarity_batch_prenorm(
     query: &[f32],
-    vectors: &[Vec<f32>],
+    vectors: &(impl VectorSet + ?Sized),
     norms: &[f32],
     tombstones: &[u8],
 ) -> Vec<(usize, f32)> {
@@ -77,7 +116,7 @@ pub fn cosine_similarity_batch_prenorm(
         return Vec::new();
     }
 
-    let n = vectors.len();
+    let n = vectors.count();
     let mut results: Vec<(usize, f32)> = Vec::with_capacity(n);
 
     for i in 0..n {
@@ -85,7 +124,7 @@ pub fn cosine_similarity_batch_prenorm(
             continue;
         }
         let vec_norm = norms[i];
-        let score = crate::cosine_similarity_prenorm(query, query_norm, &vectors[i], vec_norm);
+        let score = crate::cosine_similarity_prenorm(query, query_norm, vectors.row(i), vec_norm);
         results.push((i, score));
     }
 
@@ -162,7 +201,7 @@ pub fn cosine_similarity_f16(
 #[cfg(feature = "parallel")]
 pub fn parallel_cosine_batch(
     query: &[f32],
-    vectors: &[Vec<f32>],
+    vectors: &(impl VectorSet + Sync + ?Sized),
     tombstones: &[u8],
     k: usize,
 ) -> Vec<(usize, f32)> {
@@ -174,24 +213,27 @@ pub fn parallel_cosine_batch(
     }
 
     let num_cores = rayon::current_num_threads().max(1);
-    let chunk_size = vectors.len().div_ceil(num_cores);
+    let chunk_size = vectors.count().div_ceil(num_cores);
     if chunk_size == 0 {
         return Vec::new();
     }
 
-    let mut all_results: Vec<(usize, f32)> = vectors
-        .par_chunks(chunk_size)
-        .enumerate()
-        .flat_map(|(chunk_idx, chunk)| {
+    // Chunk over index ranges: the corpus may be one flat buffer rather than
+    // a slice of rows, so there is nothing to `par_chunks` over.
+    let n = vectors.count();
+    let mut all_results: Vec<(usize, f32)> = (0..n.div_ceil(chunk_size))
+        .into_par_iter()
+        .flat_map(|chunk_idx| {
             let base = chunk_idx * chunk_size;
-            let mut local: Vec<(usize, f32)> = Vec::with_capacity(chunk.len());
-            for (j, vec) in chunk.iter().enumerate() {
-                let i = base + j;
+            let end = (base + chunk_size).min(n);
+            let mut local: Vec<(usize, f32)> = Vec::with_capacity(end - base);
+            for i in base..end {
                 if i < tombstones.len() && tombstones[i] != 0 {
                     continue;
                 }
-                let vec_norm = clawhdf5_accel::vector_norm(vec);
-                let score = crate::cosine_similarity_prenorm(query, query_norm, vec, vec_norm);
+                let vec_norm = clawhdf5_accel::vector_norm(vectors.row(i));
+                let score =
+                    crate::cosine_similarity_prenorm(query, query_norm, vectors.row(i), vec_norm);
                 local.push((i, score));
             }
             local.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap_or(std::cmp::Ordering::Equal));
@@ -209,7 +251,7 @@ pub fn parallel_cosine_batch(
 #[cfg(feature = "parallel")]
 pub fn parallel_cosine_batch_prenorm(
     query: &[f32],
-    vectors: &[Vec<f32>],
+    vectors: &(impl VectorSet + Sync + ?Sized),
     norms: &[f32],
     tombstones: &[u8],
     k: usize,
@@ -222,23 +264,26 @@ pub fn parallel_cosine_batch_prenorm(
     }
 
     let num_cores = rayon::current_num_threads().max(1);
-    let chunk_size = vectors.len().div_ceil(num_cores);
+    let chunk_size = vectors.count().div_ceil(num_cores);
     if chunk_size == 0 {
         return Vec::new();
     }
 
-    let mut all_results: Vec<(usize, f32)> = vectors
-        .par_chunks(chunk_size)
-        .enumerate()
-        .flat_map(|(chunk_idx, chunk)| {
+    // Chunk over index ranges: the corpus may be one flat buffer rather than
+    // a slice of rows, so there is nothing to `par_chunks` over.
+    let n = vectors.count();
+    let mut all_results: Vec<(usize, f32)> = (0..n.div_ceil(chunk_size))
+        .into_par_iter()
+        .flat_map(|chunk_idx| {
             let base = chunk_idx * chunk_size;
-            let mut local: Vec<(usize, f32)> = Vec::with_capacity(chunk.len());
-            for (j, vec) in chunk.iter().enumerate() {
-                let i = base + j;
+            let end = (base + chunk_size).min(n);
+            let mut local: Vec<(usize, f32)> = Vec::with_capacity(end - base);
+            for i in base..end {
                 if i < tombstones.len() && tombstones[i] != 0 {
                     continue;
                 }
-                let score = crate::cosine_similarity_prenorm(query, query_norm, vec, norms[i]);
+                let score =
+                    crate::cosine_similarity_prenorm(query, query_norm, vectors.row(i), norms[i]);
                 local.push((i, score));
             }
             local.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap_or(std::cmp::Ordering::Equal));

crates/clawhdf5-agent/tests/float16_store.rs

@@ -0,0 +1,300 @@
+//! `MemoryConfig::float16`: embeddings stored as IEEE half precision.
+//!
+//! The setting used to be recorded in `/meta` and otherwise ignored — the
+//! embeddings dataset was always `f32`. These tests pin what it now does: the
+//! dataset is `float16`, the in-memory cache holds exactly the values the file
+//! holds (so search results survive a reopen bit for bit), and a value half
+//! precision cannot represent is refused rather than stored as infinity.
+
+use std::path::{Path, PathBuf};
+
+use clawhdf5_agent::{AgentMemory, HDF5Memory, MemoryConfig, MemoryEntry, MemoryError};
+use clawhdf5_format::float16::round_to_f16;
+use tempfile::TempDir;
+
+const DIM: usize = 64;
+
+/// Deterministic, embedding-like unit vectors.
+fn embedding(seed: u64) -> Vec<f32> {
+    let mut x = seed.wrapping_mul(0x9E37_79B9_7F4A_7C15) | 1;
+    let v: Vec<f32> = (0..DIM)
+        .map(|_| {
+            x ^= x << 13;
+            x ^= x >> 7;
+            x ^= x << 17;
+            (x >> 40) as f32 / (1u64 << 24) as f32 - 0.5
+        })
+        .collect();
+    let norm = v.iter().map(|a| a * a).sum::<f32>().sqrt();
+    v.iter().map(|a| a / norm).collect()
+}
+
+fn entry(i: u64) -> MemoryEntry {
+    MemoryEntry {
+        chunk: format!("memory number {i} about topic {}", i % 7),
+        embedding: embedding(i),
+        source_channel: "test".into(),
+        timestamp: i as f64,
+        session_id: "s".into(),
+        tags: format!("t{i}"),
+    }
+}
+
+fn config(dir: &TempDir, name: &str, float16: bool) -> MemoryConfig {
+    let mut c = MemoryConfig::new(dir.path().join(name), "agent", DIM);
+    c.float16 = float16;
+    c
+}
+
+fn embeddings_dtype_and_values(path: &Path) -> (String, Vec<f32>) {
+    let file = clawhdf5::File::open(path).unwrap();
+    let ds = file.dataset("memory/embeddings").unwrap();
+    (format!("{:?}", ds.dtype().unwrap()), ds.read_f32().unwrap())
+}
+
+fn search_bits(m: &mut HDF5Memory, q: u64) -> Vec<(usize, u32)> {
+    m.hybrid_search(&embedding(q), "memory topic 3", 0.4, 0.6, 10)
+        .iter()
+        .map(|r| (r.index, r.score.to_bits()))
+        .collect()
+}
+
+#[test]
+fn float16_store_writes_half_precision_and_reopens_identically() {
+    let dir = TempDir::new().unwrap();
+    // Two identical stores. Search is not read-only (it boosts the Hebbian
+    // activation of what it returns, and checkpoints persist that), so each
+    // is queried exactly once: one live, one after a checkpoint and reopen.
+    let live_cfg = config(&dir, "live.h5", true);
+    let cfg = config(&dir, "f16.h5", true);
+    let path: PathBuf = cfg.path.clone();
+
+    let mut live = HDF5Memory::create(live_cfg).unwrap();
+    live.save_batch((0..200).map(entry).collect()).unwrap();
+    let mut m = HDF5Memory::create(cfg).unwrap();
+    m.save_batch((0..200).map(entry).collect()).unwrap();
+    drop(m);
+
+    // On disk: a genuine float16 dataset holding the rounded inputs.
+    let (dtype, values) = embeddings_dtype_and_values(&path);
+    assert_eq!(dtype, "Other(\"float16\")");
+    let expected: Vec<u32> = (0..200)
+        .flat_map(|i| embedding(i).into_iter().map(|v| round_to_f16(v).to_bits()))
+        .collect();
+    let got: Vec<u32> = values.iter().map(|v| v.to_bits()).collect();
+    assert_eq!(got, expected);
+
+    // Reopened, the store answers exactly as the live one does: the cache
+    // held the half-rounded values before the checkpoint.
+    let mut reopened = HDF5Memory::open(&path).unwrap();
+    for q in 0..5 {
+        assert_eq!(
+            search_bits(&mut live, 1000 + q),
+            search_bits(&mut reopened, 1000 + q),
+            "query {q}"
+        );
+    }
+}
+
+#[test]
+fn float16_halves_the_embeddings_on_disk() {
+    let dir = TempDir::new().unwrap();
+    let mut sizes = Vec::new();
+    for float16 in [false, true] {
+        let cfg = config(&dir, &format!("s{float16}.h5"), float16);
+        let path = cfg.path.clone();
+        let mut m = HDF5Memory::create(cfg).unwrap();
+        m.save_batch((0..2000).map(entry).collect()).unwrap();
+        drop(m);
+        sizes.push(std::fs::metadata(&path).unwrap().len());
+    }
+    let embedding_bytes_f32 = (2000 * DIM * 4) as u64;
+    let saved = sizes[0] - sizes[1];
+    // Half of the f32 embeddings, give or take metadata and alignment.
+    assert!(
+        saved.abs_diff(embedding_bytes_f32 / 2) < 16 * 1024,
+        "f32 {} B, f16 {} B, saved {saved} B, expected ~{} B",
+        sizes[0],
+        sizes[1],
+        embedding_bytes_f32 / 2
+    );
+}
+
+#[test]
+fn f32_store_is_unchanged() {
+    let dir = TempDir::new().unwrap();
+    let cfg = config(&dir, "f32.h5", false);
+    let path = cfg.path.clone();
+    let mut m = HDF5Memory::create(cfg).unwrap();
+    m.save_batch((0..50).map(entry).collect()).unwrap();
+    drop(m);
+    let (dtype, values) = embeddings_dtype_and_values(&path);
+    assert_eq!(dtype, "F32");
+    let expected: Vec<f32> = (0..50).flat_map(embedding).collect();
+    assert_eq!(values, expected);
+}
+
+#[test]
+fn out_of_range_values_are_refused_not_stored_as_infinity() {
+    let dir = TempDir::new().unwrap();
+    let mut cfg = config(&dir, "range.h5", true);
+    cfg.wal_enabled = true;
+    let path = cfg.path.clone();
+    let mut m = HDF5Memory::create(cfg).unwrap();
+    m.save(entry(1)).unwrap();
+
+    let mut bad = entry(2);
+    bad.embedding[5] = 70_000.0;
+    match m.save(bad.clone()) {
+        Err(MemoryError::InvalidEntry(msg)) => assert!(msg.contains("embedding[5]"), "{msg}"),
+        other => panic!("expected InvalidEntry, got {other:?}"),
+    }
+    assert!(matches!(
+        m.save_or_update(bad.clone()),
+        Err(MemoryError::InvalidEntry(_))
+    ));
+    // A batch is all or nothing.
+    assert!(matches!(
+        m.save_batch(vec![entry(3), bad.clone(), entry(4)]),
+        Err(MemoryError::InvalidEntry(_))
+    ));
+    assert_eq!(m.count(), 1);
+
+    // The largest finite half, and values that round down to it, are fine.
+    let mut edge = entry(5);
+    edge.embedding[0] = 65504.0;
+    edge.embedding[1] = -65519.0;
+    m.save(edge).unwrap();
+    assert_eq!(m.count(), 2);
+    drop(m);
+
+    // Nothing rejected reached the WAL or the file.
+    let m = HDF5Memory::open(&path).unwrap();
+    assert_eq!(m.count(), 2);
+
+    // An f32 store takes the same value as it always did.
+    let mut m32 = HDF5Memory::create(config(&dir, "range32.h5", false)).unwrap();
+    m32.save(bad).unwrap();
+}
+
+#[test]
+fn wal_replay_rounds_like_a_live_save() {
+    let dir = TempDir::new().unwrap();
+    let mut cfg = config(&dir, "wal.h5", true);
+    cfg.wal_enabled = true;
+    cfg.wal_max_entries = 10_000; // keep everything in the WAL
+    let path = cfg.path.clone();
+    let mut m = HDF5Memory::create(cfg).unwrap();
+    for i in 0..30 {
+        m.save(entry(i)).unwrap();
+    }
+    let live = search_bits(&mut m, 77);
+
+    // Crash image: the .h5 is still the empty checkpoint; everything is in
+    // the WAL, which holds the caller's f32 values.
+    let crash = TempDir::new().unwrap();
+    let image = crash.path().join("image.h5");
+    std::fs::copy(&path, &image).unwrap();
+    std::fs::copy(
+        path.with_extension("h5.wal"),
+        image.with_extension("h5.wal"),
+    )
+    .unwrap();
+    drop(m);
+
+    let mut recovered = HDF5Memory::open(&image).unwrap();
+    assert_eq!(recovered.count(), 30);
+    assert_eq!(search_bits(&mut recovered, 77), live);
+}
+
+#[test]
+fn new_stores_default_to_float16() {
+    let dir = TempDir::new().unwrap();
+    let path = dir.path().join("default.h5");
+    let mut m = HDF5Memory::create(MemoryConfig::new(path.clone(), "agent", DIM)).unwrap();
+    assert!(m.config().float16);
+    m.save_batch((0..10).map(entry).collect()).unwrap();
+    drop(m);
+    assert_eq!(embeddings_dtype_and_values(&path).0, "Other(\"float16\")");
+    assert!(HDF5Memory::open(&path).unwrap().config().float16);
+}
+
+#[test]
+fn an_existing_f32_store_stays_f32() {
+    // Written by the v2.5.0 CLI, with `float16 = 0` in /meta (every agent
+    // store has recorded it). Flipping the default for new stores must not
+    // reach back and round an existing store's embeddings.
+    let dir = TempDir::new().unwrap();
+    let path = dir.path().join("legacy.h5");
+    std::fs::copy(
+        concat!(
+            env!("CARGO_MANIFEST_DIR"),
+            "/tests/fixtures/store_v2_5_0.h5"
+        ),
+        &path,
+    )
+    .unwrap();
+    let before = embeddings_dtype_and_values(&path);
+    assert_eq!(before.0, "F32");
+
+    let mut m = HDF5Memory::open(&path).unwrap();
+    assert!(!m.config().float16, "an old store must reopen as f32");
+    let dim = m.config().embedding_dim;
+    let odd: Vec<f32> = (0..dim).map(|i| 0.1 + i as f32 * 1e-4).collect();
+    m.save_batch(vec![MemoryEntry {
+        chunk: "added after the upgrade".into(),
+        embedding: odd.clone(),
+        source_channel: "test".into(),
+        timestamp: 1.0,
+        session_id: "s".into(),
+        tags: String::new(),
+    }])
+    .unwrap();
+    drop(m);
+
+    // Checkpointed: still f32, the old rows untouched and the new one exact.
+    let (dtype, values) = embeddings_dtype_and_values(&path);
+    assert_eq!(dtype, "F32");
+    assert_eq!(&values[..before.1.len()], before.1.as_slice());
+    assert_eq!(&values[before.1.len()..], odd.as_slice());
+}
+
+/// `Group::attrs` leaves out an attribute it cannot decode. A store whose
+/// `float16` setting is unreadable must not open as `float16 = false` (or with
+/// any other default in place of a setting it has): it is an error.
+#[test]
+fn unreadable_meta_attribute_fails_open_instead_of_defaulting() {
+    let dir = TempDir::new().unwrap();
+    let path = dir.path().join("store.h5");
+    {
+        let mut m = HDF5Memory::create(config(&dir, "store.h5", true)).unwrap();
+        m.save(entry(1)).unwrap();
+        m.flush_wal().unwrap();
+    }
+    assert!(HDF5Memory::open_read_only(&path).is_ok());
+
+    // Give the `float16` attribute message an unknown version (the name is
+    // at +8 in a version-1 message and +9 in a version-3 one).
+    let mut bytes = std::fs::read(&path).unwrap();
+    let name = b"float16\0";
+    let mut hit = false;
+    let positions: Vec<usize> = (9..bytes.len() - name.len())
+        .filter(|&p| &bytes[p..p + name.len()] == name)
+        .collect();
+    for pos in positions {
+        for (back, version) in [(8, 1u8), (9, 3u8)] {
+            if bytes[pos - back] == version {
+                bytes[pos - back] = 0x7f;
+                hit = true;
+            }
+        }
+    }
+    assert!(hit, "float16 attribute message not found");
+    std::fs::write(&path, &bytes).unwrap();
+
+    match HDF5Memory::open_read_only(&path) {
+        Err(MemoryError::Schema(msg)) => assert!(msg.contains("/meta"), "{msg}"),
+        Err(e) => panic!("unexpected error: {e}"),
+        Ok(_) => panic!("store opened with an unreadable float16 setting"),
+    }
+}

crates/clawhdf5-agent/tests/h5py_interop.rs

@@ -0,0 +1,155 @@
+//! An agent store is a standard HDF5 file: h5py can open it and read every
+//! dataset.
+//!
+//! It could not: the float datatype's sign-bit position was hard-coded for
+//! f64, so every f32 dataset (embeddings, norms, activation weights) made
+//! libhdf5 refuse the file with "sign bit position out of bounds".
+
+use std::process::Command;
+
+use clawhdf5_agent::{AgentMemory, HDF5Memory, MemoryConfig, MemoryEntry};
+
+fn python() -> String {
+    std::env::var("CLAWHDF5_PYTHON").unwrap_or_else(|_| "python3".to_string())
+}
+
+fn h5py_available() -> bool {
+    Command::new(python())
+        .args(["-c", "import h5py"])
+        .output()
+        .map(|o| o.status.success())
+        .unwrap_or(false)
+}
+
+#[test]
+fn h5py_reads_every_dataset_of_an_agent_store() {
+    if !h5py_available() {
+        assert!(
+            std::env::var("CLAWHDF5_REQUIRE_INTEROP").as_deref() != Ok("1"),
+            "CLAWHDF5_REQUIRE_INTEROP=1 but python3 with h5py is not available"
+        );
+        eprintln!("SKIP: python3 with h5py not available");
+        return;
+    }
+    let dir = tempfile::tempdir().unwrap();
+    for float16 in [false, true] {
+        let path = dir.path().join(format!("store_{float16}.h5"));
+        let mut cfg = MemoryConfig::new(path.clone(), "agent", 8);
+        cfg.float16 = float16;
+        let mut m = HDF5Memory::create(cfg).unwrap();
+        // save_batch checkpoints, so the records are in the .h5, not the WAL.
+        m.save_batch(
+            (0..20)
+                .map(|i| MemoryEntry {
+                    chunk: format!("memory {i}"),
+                    embedding: (0..8).map(|j| ((i * 8 + j) as f32).sin()).collect(),
+                    source_channel: "test".into(),
+                    timestamp: i as f64,
+                    session_id: "s".into(),
+                    tags: String::new(),
+                })
+                .collect(),
+        )
+        .unwrap();
+        drop(m);
+
+        // Exact expected values, as bits: numpy's sin need not match Rust's
+        // to the last place.
+        let bits = (0..160)
+            .map(|k| (k as f32).sin().to_bits().to_string())
+            .collect::<Vec<_>>()
+            .join(",");
+        let script = format!(
+            r#"
+import h5py, numpy as np
+want = np.float16 if {py_bool} else np.float32
+with h5py.File("{path}", "r") as f:
+    names = []
+    f.visititems(lambda n, o: names.append(n) if isinstance(o, h5py.Dataset) else None)
+    for n in names:
+        f[n][()]  # every dataset must decode
+    e = f["memory/embeddings"]
+    assert e.dtype == want, e.dtype
+    assert e.shape == (20, 8), e.shape
+    ref = np.array([{bits}], dtype=np.uint32).view(np.float32).astype(want).reshape(20, 8)
+    assert (e[()] == ref).all()
+    assert f["memory/norms"].dtype == np.float32
+print(len(names))
+"#,
+            py_bool = if float16 { "True" } else { "False" },
+            path = path.display()
+        );
+        let out = Command::new(python())
+            .args(["-c", &script])
+            .output()
+            .unwrap();
+        assert!(
+            out.status.success(),
+            "float16={float16}: {}",
+            String::from_utf8_lossy(&out.stderr)
+        );
+        let n: usize = String::from_utf8_lossy(&out.stdout).trim().parse().unwrap();
+        assert!(n >= 10, "only {n} datasets");
+    }
+}
+
+#[test]
+fn an_edit_made_with_h5py_breaks_the_signature_and_names_the_record() {
+    if !h5py_available() {
+        assert!(
+            std::env::var("CLAWHDF5_REQUIRE_INTEROP").as_deref() != Ok("1"),
+            "CLAWHDF5_REQUIRE_INTEROP=1 but python3 with h5py is not available"
+        );
+        eprintln!("SKIP: python3 with h5py not available");
+        return;
+    }
+    use clawhdf5_agent::signing::SigningKey;
+    let dir = tempfile::tempdir().unwrap();
+    let path = dir.path().join("signed.h5");
+    let key = SigningKey::from_bytes(&[42; 32]);
+    let mut m = HDF5Memory::create(MemoryConfig::new(path.clone(), "agent", 8)).unwrap();
+    m.set_signing_key(key.clone());
+    m.save_batch(
+        (0..10)
+            .map(|i| MemoryEntry {
+                chunk: format!("memory {i}"),
+                embedding: (0..8).map(|j| ((i * 8 + j) as f32).cos()).collect(),
+                source_channel: "test".into(),
+                timestamp: i as f64,
+                session_id: "s".into(),
+                tags: String::new(),
+            })
+            .collect(),
+    )
+    .unwrap();
+    drop(m);
+    assert!(
+        HDF5Memory::verify(&path, &key.verifying_key())
+            .unwrap()
+            .is_valid()
+    );
+
+    // Someone edits one timestamp in place with h5py.
+    let script = format!(
+        r#"
+import h5py
+with h5py.File("{}", "r+") as f:
+    ts = f["memory/timestamps"]
+    ts[3] = 12345.0
+"#,
+        path.display()
+    );
+    let out = Command::new(python())
+        .args(["-c", &script])
+        .output()
+        .unwrap();
+    assert!(
+        out.status.success(),
+        "{}",
+        String::from_utf8_lossy(&out.stderr)
+    );
+
+    let r = HDF5Memory::verify(&path, &key.verifying_key()).unwrap();
+    assert!(r.signature_valid && !r.is_valid(), "{r:?}");
+    assert_eq!(r.changed_records, vec![3]);
+}

crates/clawhdf5-agent/tests/hnsw_integration.rs

@@ -165,3 +165,182 @@ fn save_batch_then_search_is_consistent() {
         );
     }
 }
+
+#[test]
+fn quantized_index_matches_the_f32_index_after_re_scoring() {
+    // A quantised index holds approximate vectors, but the store still has the
+    // exact ones, so the query path re-scores the candidate pool before
+    // fusion. The results a caller sees should therefore be the same.
+    let dim = 64;
+    let n = 400;
+    let mut seed = 0x5EED_1234_5678_9ABC;
+    let vectors: Vec<Vec<f32>> = (0..n).map(|_| make_vector(&mut seed, dim)).collect();
+    let queries: Vec<Vec<f32>> = (0..20).map(|_| make_vector(&mut seed, dim)).collect();
+
+    let build = |dir: &TempDir, quantized: bool| {
+        let mut config = MemoryConfig::new(dir.path().join("mem.h5"), "agent", dim);
+        config.quantized_index = quantized;
+        let mut mem = HDF5Memory::create(config).unwrap();
+        for (i, v) in vectors.iter().enumerate() {
+            mem.save(entry(&format!("chunk {i}"), v.clone(), &format!("k{i}")))
+                .unwrap();
+        }
+        mem
+    };
+
+    let exact_dir = TempDir::new().unwrap();
+    let quant_dir = TempDir::new().unwrap();
+    let mut exact = build(&exact_dir, false);
+    let mut quantized = build(&quant_dir, true);
+
+    let k = 10;
+    let mut agree = 0;
+    for q in &queries {
+        let want: Vec<usize> = exact
+            .hybrid_search(q, "", 1.0, 0.0, k)
+            .iter()
+            .map(|r| r.index)
+            .collect();
+        agree += quantized
+            .hybrid_search(q, "", 1.0, 0.0, k)
+            .iter()
+            .filter(|r| want.contains(&r.index))
+            .count();
+    }
+    let overlap = agree as f64 / (k * queries.len()) as f64;
+    assert!(
+        overlap >= 0.95,
+        "quantised store should match the f32 one: {overlap}"
+    );
+}
+
+#[test]
+fn quantized_index_setting_survives_a_reopen() {
+    let dir = TempDir::new().unwrap();
+    let path = dir.path().join("mem.h5");
+    let mut config = MemoryConfig::new(path.clone(), "agent", 8);
+    config.quantized_index = true;
+    let mut mem = HDF5Memory::create(config).unwrap();
+    let mut seed = 7;
+    for i in 0..30 {
+        mem.save(entry(&format!("c{i}"), make_vector(&mut seed, 8), "t"))
+            .unwrap();
+    }
+    mem.flush_wal().unwrap();
+    drop(mem);
+
+    // Reopening must not silently quadruple the index's memory, so the flag
+    // is part of the stored config rather than a per-session choice.
+    let reopened = HDF5Memory::open(&path).unwrap();
+    assert!(reopened.config().quantized_index);
+}
+
+#[test]
+fn hnsw_parameters_are_configurable_and_persisted() {
+    // The graph degree and both candidate-list sizes used to be constants, so
+    // a deployment could not trade recall against memory or speed at all.
+    let dir = TempDir::new().unwrap();
+    let path = dir.path().join("mem.h5");
+    let mut config = MemoryConfig::new(path.clone(), "agent", 16);
+    config.hnsw_m = 8;
+    config.hnsw_ef_construction = 32;
+    config.hnsw_ef_search = 128;
+    let mut mem = HDF5Memory::create(config).unwrap();
+
+    let mut seed = 99;
+    let vectors: Vec<Vec<f32>> = (0..300).map(|_| make_vector(&mut seed, 16)).collect();
+    for (i, v) in vectors.iter().enumerate() {
+        mem.save(entry(&format!("c{i}"), v.clone(), "t")).unwrap();
+    }
+    // Still correct with a smaller graph: an exact match must rank first.
+    let top = mem.hybrid_search(&vectors[42], "", 1.0, 0.0, 1);
+    assert_eq!(top[0].index, 42);
+
+    mem.flush_wal().unwrap();
+    drop(mem);
+    let reopened = HDF5Memory::open(&path).unwrap();
+    assert_eq!(reopened.config().hnsw_m, 8);
+    assert_eq!(reopened.config().hnsw_ef_construction, 32);
+    assert_eq!(reopened.config().hnsw_ef_search, 128);
+}
+
+#[test]
+fn degenerate_hnsw_parameters_do_not_panic() {
+    // `clawhdf5-ann` asserts m >= 2, so a zero from a config file — or from a
+    // caller who assumed 0 meant "default" — would abort the process inside
+    // the index builder. The store clamps instead.
+    let dir = TempDir::new().unwrap();
+    let mut config = MemoryConfig::new(dir.path().join("mem.h5"), "agent", 8);
+    config.hnsw_m = 0;
+    config.hnsw_ef_construction = 0;
+    config.hnsw_ef_search = 1;
+    let mut mem = HDF5Memory::create(config).unwrap();
+
+    let mut seed = 5;
+    let vectors: Vec<Vec<f32>> = (0..50).map(|_| make_vector(&mut seed, 8)).collect();
+    for (i, v) in vectors.iter().enumerate() {
+        mem.save(entry(&format!("c{i}"), v.clone(), "t")).unwrap();
+    }
+    let results = mem.hybrid_search(&vectors[7], "", 1.0, 0.0, 5);
+    assert_eq!(results[0].index, 7, "exact match should still rank first");
+}
+
+#[test]
+fn new_stores_default_to_the_quantized_index() {
+    // int8 is the default because it is smaller and, with an exact re-score,
+    // faster at equal recall on every platform measured (see BENCHMARKS.md).
+    let dir = TempDir::new().unwrap();
+    let config = MemoryConfig::new(dir.path().join("mem.h5"), "agent", 8);
+    assert!(config.quantized_index);
+
+    let path = config.path.clone();
+    let mut mem = HDF5Memory::create(config).unwrap();
+    let mut seed = 3;
+    let vectors: Vec<Vec<f32>> = (0..40).map(|_| make_vector(&mut seed, 8)).collect();
+    for (i, v) in vectors.iter().enumerate() {
+        mem.save(entry(&format!("c{i}"), v.clone(), "t")).unwrap();
+    }
+    assert_eq!(
+        mem.hybrid_search(&vectors[11], "", 1.0, 0.0, 1)[0].index,
+        11
+    );
+    mem.flush_wal().unwrap();
+    drop(mem);
+    assert!(HDF5Memory::open(&path).unwrap().config().quantized_index);
+}
+
+#[test]
+fn a_store_written_before_the_setting_existed_stays_f32() {
+    // `store_v2_5_0.h5` was written by the v2.5.0 CLI, before
+    // `quantized_index` or the HNSW parameters were persisted, so it carries
+    // none of them. Flipping the default for new stores must not reach back
+    // and change how an existing store's index is held.
+    let dir = TempDir::new().unwrap();
+    let path = dir.path().join("legacy.h5");
+    std::fs::copy(
+        concat!(
+            env!("CARGO_MANIFEST_DIR"),
+            "/tests/fixtures/store_v2_5_0.h5"
+        ),
+        &path,
+    )
+    .unwrap();
+
+    let bytes = std::fs::read(&path).unwrap();
+    assert!(
+        !bytes.windows(15).any(|w| w == b"quantized_index"),
+        "the fixture must predate the setting, or it tests nothing"
+    );
+
+    let mut mem = HDF5Memory::open(&path).unwrap();
+    assert!(
+        !mem.config().quantized_index,
+        "an old store must reopen with an f32 index"
+    );
+    assert_eq!(mem.config().hnsw_m, 16);
+    assert_eq!(mem.config().hnsw_ef_construction, 64);
+    assert_eq!(mem.count(), 6);
+    // And it still searches: entry 3's own embedding finds it first.
+    let hit = mem.hybrid_search(&[3.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], "", 1.0, 0.0, 1);
+    assert_eq!(hit[0].index, 3);
+}

crates/clawhdf5-agent/tests/search_options.rs

@@ -0,0 +1,344 @@
+//! `HDF5Memory::search` with `SearchOptions`: source filtering, re-ranking and
+//! confidence rejection in the store's own search path.
+
+use std::collections::HashSet;
+
+use clawhdf5_agent::confidence::ConfidenceConfig;
+use clawhdf5_agent::reranker::ReRankConfig;
+use clawhdf5_agent::{AgentMemory, HDF5Memory, MemoryConfig, MemoryEntry, SearchOptions, hybrid};
+use tempfile::TempDir;
+
+const DIM: usize = 32;
+const N: usize = 3000;
+const CLUSTERS: usize = 20;
+
+struct Rng(u64);
+impl Rng {
+    fn next(&mut self) -> u64 {
+        self.0 = self.0.wrapping_add(0x9E37_79B9_7F4A_7C15);
+        let mut z = self.0;
+        z = (z ^ (z >> 30)).wrapping_mul(0xBF58_476D_1CE4_E5B9);
+        z = (z ^ (z >> 27)).wrapping_mul(0x94D0_49BB_1331_11EB);
+        z ^ (z >> 31)
+    }
+    fn unit(&mut self) -> f32 {
+        (self.next() >> 40) as f32 / (1u64 << 24) as f32 - 0.5
+    }
+}
+
+fn normalize(v: &mut [f32]) {
+    let n = v.iter().map(|x| x * x).sum::<f32>().sqrt();
+    v.iter_mut().for_each(|x| *x /= n);
+}
+
+struct Data {
+    vectors: Vec<Vec<f32>>,
+    cluster: Vec<usize>,
+    centres: Vec<Vec<f32>>,
+}
+
+fn data() -> Data {
+    let mut rng = Rng(42);
+    let centres: Vec<Vec<f32>> = (0..CLUSTERS)
+        .map(|_| {
+            let mut c: Vec<f32> = (0..DIM).map(|_| rng.unit()).collect();
+            normalize(&mut c);
+            c
+        })
+        .collect();
+    let mut vectors = Vec::new();
+    let mut cluster = Vec::new();
+    for i in 0..N {
+        let c = i % CLUSTERS;
+        let mut v: Vec<f32> = centres[c].iter().map(|x| x + rng.unit() * 0.3).collect();
+        normalize(&mut v);
+        vectors.push(v);
+        cluster.push(c);
+    }
+    Data {
+        vectors,
+        cluster,
+        centres,
+    }
+}
+
+/// Channel of record `i` for a filter keeping `percent`% of the store at
+/// random (independent of the vectors).
+fn random_channel(i: usize, rng_seed: u64, percent: u64) -> String {
+    let mut r = Rng(rng_seed ^ (i as u64 * 7919));
+    if r.next() % 100 < percent {
+        "keep".into()
+    } else {
+        "other".into()
+    }
+}
+
+fn build(data: &Data, channel: impl Fn(usize) -> String) -> (TempDir, HDF5Memory) {
+    let dir = TempDir::new().unwrap();
+    let mut cfg = MemoryConfig::new(dir.path().join("s.h5"), "agent", DIM);
+    cfg.hebbian_boost = 0.0; // every query sees the same store
+    let mut m = HDF5Memory::create(cfg).unwrap();
+    let entries = data
+        .vectors
+        .iter()
+        .enumerate()
+        .map(|(i, v)| MemoryEntry {
+            chunk: format!("record {i} cluster {}", data.cluster[i]),
+            embedding: v.clone(),
+            source_channel: channel(i),
+            timestamp: i as f64,
+            session_id: "s".into(),
+            tags: format!("t{i}"),
+        })
+        .collect();
+    m.save_batch(entries).unwrap();
+    (dir, m)
+}
+
+/// Exact top-k by cosine among the records `allowed` keeps.
+fn exact_top(data: &Data, q: &[f32], k: usize, allowed: impl Fn(usize) -> bool) -> Vec<usize> {
+    let mut s: Vec<(usize, f32)> = (0..N)
+        .filter(|&i| allowed(i))
+        .map(|i| (i, data.vectors[i].iter().zip(q).map(|(a, b)| a * b).sum()))
+        .collect();
+    s.sort_by(|a, b| b.1.total_cmp(&a.1).then(a.0.cmp(&b.0)));
+    s.into_iter().take(k).map(|(i, _)| i).collect()
+}
+
+fn query(data: &Data, i: usize) -> Vec<f32> {
+    let mut rng = Rng(1000 + i as u64);
+    let mut q: Vec<f32> = data.centres[i % CLUSTERS]
+        .iter()
+        .map(|x| x + rng.unit() * 0.3)
+        .collect();
+    normalize(&mut q);
+    q
+}
+
+fn vector_only(k: usize) -> SearchOptions {
+    SearchOptions::new(k).with_fusion(hybrid::Fusion::Weighted {
+        vector: 1.0,
+        keyword: 0.0,
+    })
+}
+
+#[test]
+fn source_filter_returns_only_allowed_records_and_a_full_page() {
+    let d = data();
+    // At N = 3000 and k = 10 the index serves a filter only when that is
+    // cheaper than scanning the allowed records: pool = 80 * N / allowed
+    // candidates at ~M = 16 distances each, against `allowed` distances. So
+    // 90% goes through the index, 50% and 1% to the exact scan.
+    for percent in [90, 50, 1] {
+        let (_dir, mut m) = build(&d, |i| random_channel(i, 5, percent));
+        let allowed = |i: usize| random_channel(i, 5, percent) == "keep";
+        let mut hits = 0;
+        for qi in 0..40 {
+            let q = query(&d, qi);
+            let got = m.search(&q, "", &vector_only(10).with_sources(["keep"]));
+            assert_eq!(got.len(), 10, "{percent}%: short page");
+            assert!(got.iter().all(|r| r.source_channel == "keep"));
+            let want: HashSet<usize> = exact_top(&d, &q, 10, allowed).into_iter().collect();
+            hits += got.iter().filter(|r| want.contains(&r.index)).count();
+        }
+        let recall = hits as f64 / 400.0;
+        let floor = if percent == 90 { 0.95 } else { 1.0 };
+        assert!(recall >= floor, "{percent}%: recall@10 {recall}");
+    }
+}
+
+#[test]
+fn filter_away_from_the_query_falls_back_to_an_exact_scan() {
+    // Channel = cluster, and the filter keeps two clusters (10% of the
+    // store) that are not the query's: the index's neighbourhood of the
+    // query holds none of them. The search must still return the exact
+    // top 10 among the allowed records, not a short or empty page.
+    let d = data();
+    let (_dir, mut m) = build(&d, |i| format!("c{}", d.cluster[i]));
+    for qi in 0..20 {
+        let q = query(&d, qi);
+        let a = format!("c{}", (qi + 7) % CLUSTERS);
+        let b = format!("c{}", (qi + 13) % CLUSTERS);
+        let got: Vec<usize> = m
+            .search(
+                &q,
+                "",
+                &vector_only(10).with_sources([a.clone(), b.clone()]),
+            )
+            .iter()
+            .map(|r| r.index)
+            .collect();
+        let want = exact_top(&d, &q, 10, |i| {
+            let c = format!("c{}", d.cluster[i]);
+            c == a || c == b
+        });
+        assert_eq!(got, want, "query {qi}");
+    }
+}
+
+#[test]
+fn filter_edge_cases() {
+    let d = data();
+    let (_dir, mut m) = build(&d, |i| random_channel(i, 9, 50));
+    let q = query(&d, 0);
+    assert!(
+        m.search(
+            &q,
+            "cluster",
+            &SearchOptions::new(10).with_sources(Vec::<String>::new())
+        )
+        .is_empty()
+    );
+    assert!(
+        m.search(
+            &q,
+            "cluster",
+            &SearchOptions::new(10).with_sources(["nope"])
+        )
+        .is_empty()
+    );
+    // Keyword matches from other channels are filtered too.
+    let got = m.search(
+        &q,
+        "record cluster",
+        &SearchOptions::new(50).with_sources(["keep"]),
+    );
+    assert_eq!(got.len(), 50);
+    assert!(got.iter().all(|r| r.source_channel == "keep"));
+    // Deleted records never come back, filtered or not.
+    let first = got[0].index;
+    m.delete(first).unwrap();
+    let again = m.search(
+        &q,
+        "record cluster",
+        &SearchOptions::new(50).with_sources(["keep"]),
+    );
+    assert!(again.iter().all(|r| r.index != first));
+}
+
+#[test]
+fn plain_options_equal_hybrid_search_with() {
+    // Two identical stores, so neither query sees the other's boosts.
+    let d = data();
+    let (_a, mut a) = build(&d, |i| random_channel(i, 3, 50));
+    let (_b, mut b) = build(&d, |i| random_channel(i, 3, 50));
+    for qi in 0..10 {
+        let q = query(&d, qi);
+        let x: Vec<(usize, u32)> = a
+            .search(&q, "record cluster 3", &SearchOptions::new(10))
+            .iter()
+            .map(|r| (r.index, r.score.to_bits()))
+            .collect();
+        let y: Vec<(usize, u32)> = b
+            .hybrid_search_with(&q, "record cluster 3", hybrid::DEFAULT_FUSION, 10)
+            .iter()
+            .map(|r| (r.index, r.score.to_bits()))
+            .collect();
+        assert_eq!(x, y);
+    }
+}
+
+fn small_store(entries: &[(&str, &str, f64)]) -> (TempDir, HDF5Memory) {
+    let dir = TempDir::new().unwrap();
+    let mut m = HDF5Memory::create(MemoryConfig::new(dir.path().join("r.h5"), "a", 4)).unwrap();
+    m.save_batch(
+        entries
+            .iter()
+            .map(|(chunk, channel, ts)| MemoryEntry {
+                chunk: chunk.to_string(),
+                embedding: vec![1.0, 0.0, 0.0, 0.0],
+                source_channel: channel.to_string(),
+                timestamp: *ts,
+                session_id: "s".into(),
+                tags: String::new(),
+            })
+            .collect(),
+    )
+    .unwrap();
+    (dir, m)
+}
+
+#[test]
+fn rerank_breaks_relevance_ties_by_recency() {
+    // Identical text and vectors, so retrieval ties; re-ranking must put the
+    // newer record first and report the combined score.
+    let now = 1_000_000.0;
+    let (_d, mut m) = small_store(&[
+        ("user prefers dark mode", "chat", now - 30.0 * 86_400.0),
+        ("user prefers dark mode", "chat", now - 60.0),
+    ]);
+    let q = [1.0, 0.0, 0.0, 0.0];
+    let plain = m.search(&q, "dark mode", &SearchOptions::new(2));
+    assert_eq!(plain[0].index, 0, "ties break by index without re-ranking");
+    let reranked = m.search(
+        &q,
+        "dark mode",
+        &SearchOptions::new(2)
+            .with_rerank(ReRankConfig::default())
+            .at_time(now),
+    );
+    assert_eq!(reranked[0].index, 1);
+    assert!(reranked[0].score > reranked[1].score);
+    assert_ne!(reranked[0].score.to_bits(), plain[0].score.to_bits());
+}
+
+#[test]
+fn confidence_rejects_when_nothing_is_good_enough() {
+    let (_d, mut m) = small_store(&[("alpha", "chat", 0.0), ("beta", "chat", 0.0)]);
+    let q = [1.0, 0.0, 0.0, 0.0];
+    let strict = ConfidenceConfig {
+        min_score: 10.0,
+        ..ConfidenceConfig::default()
+    };
+    assert!(
+        m.search(&q, "alpha", &SearchOptions::new(2).with_confidence(strict))
+            .is_empty()
+    );
+    let lenient = ConfidenceConfig {
+        min_score: 0.0,
+        min_gap: f32::INFINITY,
+        max_results: 1,
+    };
+    assert_eq!(
+        m.search(&q, "alpha", &SearchOptions::new(2).with_confidence(lenient))
+            .len(),
+        1
+    );
+}
+
+#[test]
+fn only_returned_results_are_reinforced() {
+    // With re-ranking, a pool of max(3k, 10) candidates is retrieved; only
+    // the k returned should gain activation.
+    let d = data();
+    let dir = TempDir::new().unwrap();
+    let path = dir.path().join("h.h5");
+    let mut m = HDF5Memory::create(MemoryConfig::new(path, "a", DIM)).unwrap();
+    m.save_batch(
+        (0..200)
+            .map(|i| MemoryEntry {
+                chunk: format!("record {i}"),
+                embedding: d.vectors[i].clone(),
+                source_channel: "chat".into(),
+                timestamp: i as f64,
+                session_id: "s".into(),
+                tags: String::new(),
+            })
+            .collect(),
+    )
+    .unwrap();
+    let q = query(&d, 0);
+    let got = m.search(
+        &q,
+        "record",
+        &SearchOptions::new(3).with_rerank(ReRankConfig::default()),
+    );
+    assert_eq!(got.len(), 3);
+    let returned: HashSet<usize> = got.iter().map(|r| r.index).collect();
+    // A second plain search reports each record's current activation.
+    let all = m.search(&q, "record", &SearchOptions::new(200));
+    for r in &all {
+        let boosted = r.activation > 1.0;
+        assert_eq!(boosted, returned.contains(&r.index), "record {}", r.index);
+    }
+}

crates/clawhdf5-agent/tests/signed_store.rs

@@ -0,0 +1,330 @@
+//! Ed25519-signed checkpoints: `HDF5Memory::set_signing_key` and
+//! `HDF5Memory::verify`.
+
+use std::path::Path;
+
+use clawhdf5_agent::signing::{SigningKey, VerifyReport, VerifyingKey};
+use clawhdf5_agent::storage;
+use clawhdf5_agent::{AgentMemory, HDF5Memory, MemoryConfig, MemoryEntry, MemoryError, schema};
+use tempfile::TempDir;
+
+const DIM: usize = 16;
+
+fn key(seed: u8) -> SigningKey {
+    SigningKey::from_bytes(&[seed; 32])
+}
+
+fn entry(i: usize, chunk: &str) -> MemoryEntry {
+    MemoryEntry {
+        chunk: chunk.to_string(),
+        embedding: (0..DIM)
+            .map(|j| ((i * DIM + j) as f32 * 0.37).sin())
+            .collect(),
+        source_channel: "chat".into(),
+        timestamp: 1_700_000_000.0 + i as f64,
+        session_id: format!("s{}", i % 3),
+        tags: format!("t{i}"),
+    }
+}
+
+/// Awkward strings on purpose: they must hash the same after a round trip.
+const TEXTS: [&str; 6] = [
+    "plain text",
+    "ünïcödé — 日本語 🙂",
+    "",
+    "trailing spaces   ",
+    "tab\tand\nnewline",
+    "x",
+];
+
+fn signed_store(dir: &TempDir, float16: bool, k: &SigningKey) -> std::path::PathBuf {
+    let mut cfg = MemoryConfig::new(dir.path().join("s.h5"), "agent", DIM);
+    cfg.float16 = float16;
+    let path = cfg.path.clone();
+    let mut m = HDF5Memory::create(cfg).unwrap();
+    m.set_signing_key(k.clone());
+    let entries = (0..30).map(|i| entry(i, TEXTS[i % TEXTS.len()])).collect();
+    m.save_batch(entries).unwrap();
+    // Some graph and a deleted record, so every part of the manifest is used.
+    let a = m.knowledge_mut().add_entity("Alice", "person", 0);
+    let b = m.knowledge_mut().add_entity("Acme", "org", -1);
+    m.knowledge_mut().add_relation(a, b, "works_at", 0.75);
+    m.sessions_mut()
+        .add_at("s0", 0, 9, "chat", "first session", 1_700_000_000.0);
+    m.delete(4).unwrap();
+    m.flush_wal().unwrap();
+    path
+}
+
+fn verify(path: &Path, k: &SigningKey) -> VerifyReport {
+    HDF5Memory::verify(path, &k.verifying_key()).unwrap()
+}
+
+#[test]
+fn a_signed_store_verifies_through_reopen_and_checkpoint_cycles() {
+    for float16 in [true, false] {
+        let dir = TempDir::new().unwrap();
+        let k = key(7);
+        let path = signed_store(&dir, float16, &k);
+        let r = verify(&path, &k);
+        assert!(r.is_valid(), "float16={float16}: {r:?}");
+        assert_eq!(r.public_key, Some(k.verifying_key().to_bytes()));
+        assert_eq!(r.record_count, 30);
+        assert!(r.changed_records.is_empty());
+
+        // Reopen, change nothing, checkpoint again (with the key): still valid.
+        for _ in 0..3 {
+            let mut m = HDF5Memory::open(&path).unwrap();
+            assert!(m.is_signed());
+            m.set_signing_key(k.clone());
+            m.flush_wal().unwrap();
+            drop(m);
+            assert!(verify(&path, &k).is_valid());
+        }
+        // And after real changes, re-signed.
+        let mut m = HDF5Memory::open(&path).unwrap();
+        m.set_signing_key(k.clone());
+        m.save(entry(99, "added later")).unwrap();
+        m.hybrid_search(&entry(1, "").embedding, "text", 0.4, 0.6, 5);
+        m.flush_wal().unwrap();
+        drop(m);
+        let r = verify(&path, &k);
+        assert!(r.is_valid(), "{r:?}");
+        assert_eq!(r.record_count, 31);
+    }
+}
+
+#[test]
+fn a_signed_store_refuses_to_checkpoint_without_its_key() {
+    let dir = TempDir::new().unwrap();
+    let k = key(1);
+    let path = signed_store(&dir, true, &k);
+
+    let mut m = HDF5Memory::open(&path).unwrap();
+    m.save(entry(50, "pending")).unwrap();
+    match m.flush_wal() {
+        Err(MemoryError::SigningKeyRequired(msg)) => assert!(msg.contains("signed"), "{msg}"),
+        other => panic!("expected SigningKeyRequired, got {other:?}"),
+    }
+    // The file is untouched and still valid; the save is still in the WAL.
+    let r = verify(&path, &k);
+    assert!(r.is_valid());
+    assert_eq!(r.wal_entries_unsigned, 1);
+
+    // Supplying the key lets the checkpoint through, signed.
+    m.set_signing_key(k.clone());
+    m.flush_wal().unwrap();
+    drop(m);
+    let r = verify(&path, &k);
+    assert!(r.is_valid());
+    assert_eq!((r.record_count, r.wal_entries_unsigned), (31, 0));
+
+    // Removing the signature on purpose writes it unsigned.
+    let mut m = HDF5Memory::open(&path).unwrap();
+    m.remove_signature();
+    m.flush_wal().unwrap();
+    drop(m);
+    let r = verify(&path, &k);
+    assert!(!r.signed && !r.is_valid());
+    assert!(!HDF5Memory::open(&path).unwrap().is_signed());
+}
+
+#[test]
+fn the_wrong_key_does_not_verify_and_a_new_key_re_signs() {
+    let dir = TempDir::new().unwrap();
+    let (a, b) = (key(1), key(2));
+    let path = signed_store(&dir, true, &a);
+    let r = verify(&path, &b);
+    assert!(r.signed && !r.key_matches && !r.signature_valid && !r.is_valid());
+
+    let mut m = HDF5Memory::open(&path).unwrap();
+    m.set_signing_key(b.clone());
+    m.flush_wal().unwrap();
+    drop(m);
+    assert!(verify(&path, &b).is_valid());
+    assert!(!verify(&path, &a).is_valid());
+}
+
+/// Rewrite the store with changed contents but the *old* signature — what
+/// someone with write access to the file, but not the key, can do.
+fn tamper(path: &Path, change: impl FnOnce(&mut Tampered)) {
+    let file = clawhdf5::File::open(path).unwrap();
+    let (config, cache, sessions, knowledge) = schema::validate_and_load(&file).unwrap();
+    let checkpoint = schema::read_checkpoint_meta(&file);
+    let signature = schema::read_signature(&file).unwrap().unwrap();
+    drop(file);
+    let mut t = Tampered {
+        config,
+        cache,
+        sessions,
+        knowledge,
+    };
+    change(&mut t);
+    storage::write_to_disk_signed(
+        path,
+        &t.config,
+        &t.cache,
+        &t.sessions,
+        &t.knowledge,
+        &checkpoint,
+        Some(&signature),
+    )
+    .unwrap();
+}
+
+struct Tampered {
+    config: MemoryConfig,
+    cache: clawhdf5_agent::cache::MemoryCache,
+    sessions: clawhdf5_agent::SessionCache,
+    knowledge: clawhdf5_agent::knowledge::KnowledgeCache,
+}
+
+#[test]
+fn every_kind_of_edit_is_detected_and_located() {
+    let k = key(3);
+    type Edit = Box<dyn FnOnce(&mut Tampered)>;
+    type Case = (&'static str, Edit, fn(&VerifyReport) -> bool);
+    let cases: Vec<Case> = vec![
+        (
+            "record text",
+            Box::new(|t: &mut Tampered| t.cache.chunks[7] = "rewritten".into()),
+            |r| !r.records_match && r.changed_records == vec![7],
+        ),
+        (
+            "one embedding value",
+            Box::new(|t: &mut Tampered| {
+                let mut e = t.cache.embeddings[12].to_vec();
+                e[3] = 0.5;
+                t.cache.embeddings.set(12, &e);
+            }),
+            |r| r.changed_records == vec![12],
+        ),
+        (
+            "undelete",
+            Box::new(|t: &mut Tampered| t.cache.tombstones[4] = 0),
+            |r| r.changed_records == vec![4],
+        ),
+        (
+            "timestamp",
+            Box::new(|t: &mut Tampered| t.cache.timestamps[20] += 1.0),
+            |r| r.changed_records == vec![20],
+        ),
+        (
+            "record appended",
+            Box::new(|t: &mut Tampered| {
+                t.cache.push(
+                    "new".into(),
+                    vec![0.1; DIM],
+                    "x".into(),
+                    1.0,
+                    "s".into(),
+                    "".into(),
+                );
+            }),
+            |r| !r.records_match && r.changed_records == vec![30] && r.record_count == 31,
+        ),
+        (
+            "setting",
+            Box::new(|t: &mut Tampered| t.config.agent_id = "someone-else".into()),
+            |r| !r.settings_match && r.records_match,
+        ),
+        (
+            "session summary",
+            Box::new(|t: &mut Tampered| t.sessions.summaries[0] = "edited".into()),
+            |r| !r.sessions_match && r.records_match,
+        ),
+        (
+            "graph edge",
+            Box::new(|t: &mut Tampered| t.knowledge.relations[0].weight = 1.0),
+            |r| !r.graph_match && r.records_match,
+        ),
+    ];
+    for (name, edit, check) in cases {
+        let dir = TempDir::new().unwrap();
+        let path = signed_store(&dir, true, &k);
+        tamper(&path, edit);
+        let r = verify(&path, &k);
+        assert!(
+            r.signed && r.key_matches && r.signature_valid,
+            "{name}: {r:?}"
+        );
+        assert!(!r.is_valid(), "{name}: edit not detected: {r:?}");
+        assert!(check(&r), "{name}: {r:?}");
+    }
+}
+
+#[test]
+fn a_forged_manifest_fails_the_signature() {
+    // Recomputing the hashes for tampered contents does not help without the
+    // key: the signature no longer matches the manifest.
+    let dir = TempDir::new().unwrap();
+    let k = key(5);
+    let path = signed_store(&dir, true, &k);
+    let file = clawhdf5::File::open(&path).unwrap();
+    let (config, mut cache, sessions, knowledge) = schema::validate_and_load(&file).unwrap();
+    let checkpoint = schema::read_checkpoint_meta(&file);
+    let mut sig = schema::read_signature(&file).unwrap().unwrap();
+    drop(file);
+    cache.chunks[0] = "forged".into();
+    // Re-sign with an attacker key, then splice the victim's public key back.
+    let forged = clawhdf5_agent::signing::sign(
+        &key(66),
+        &config,
+        &cache,
+        &sessions,
+        &knowledge,
+        checkpoint.wal_applied,
+    );
+    sig.manifest = forged.manifest;
+    sig.record_hashes = forged.record_hashes;
+    storage::write_to_disk_signed(
+        &path,
+        &config,
+        &cache,
+        &sessions,
+        &knowledge,
+        &checkpoint,
+        Some(&sig),
+    )
+    .unwrap();
+    let r = verify(&path, &k);
+    assert!(
+        r.key_matches && !r.signature_valid && !r.is_valid(),
+        "{r:?}"
+    );
+}
+
+#[test]
+fn an_unsigned_store_reports_unsigned() {
+    let dir = TempDir::new().unwrap();
+    let mut m = HDF5Memory::create(MemoryConfig::new(dir.path().join("u.h5"), "a", DIM)).unwrap();
+    m.save_batch(vec![entry(0, "hello")]).unwrap();
+    drop(m);
+    let r = HDF5Memory::verify(&dir.path().join("u.h5"), &VerifyingKey::from(&key(1))).unwrap();
+    assert!(!r.signed && !r.is_valid());
+    assert_eq!(r.record_count, 1);
+}
+
+#[test]
+fn nul_bytes_in_text_still_verify() {
+    // Strings are stored null-padded; the hash must follow what a reopened
+    // store actually holds, or an untouched store would fail to verify.
+    let dir = TempDir::new().unwrap();
+    let k = key(9);
+    let mut m = HDF5Memory::create(MemoryConfig::new(dir.path().join("n.h5"), "a", DIM)).unwrap();
+    m.set_signing_key(k.clone());
+    m.save_batch(vec![
+        entry(0, "inner\0nul"),
+        entry(1, "trailing nul\0"),
+        entry(2, "\0leading"),
+    ])
+    .unwrap();
+    drop(m);
+    let r = verify(&dir.path().join("n.h5"), &k);
+    assert!(r.is_valid(), "{r:?}");
+    let m = HDF5Memory::open(&dir.path().join("n.h5")).unwrap();
+    eprintln!(
+        "reloaded: {:?}",
+        (0..3).map(|i| m.get_chunk(i)).collect::<Vec<_>>()
+    );
+}

crates/clawhdf5-android/Cargo.toml

@@ -1,8 +1,9 @@
 [package]
 name = "clawhdf5-android"
-version = "2.5.0"
+version = "2.7.0"
 edition = "2024"
-description = "Android JNI bridge for edgehdf5-memory HDF5 backend"
+rust-version.workspace = true
+description = "Android JNI bindings for clawhdf5 agent memory"
 license = "MIT"
 
 [lib]

crates/clawhdf5-android/README.md

@@ -0,0 +1,44 @@
+# clawhdf5-android
+
+A C ABI over [`clawhdf5-agent`](../clawhdf5-agent/README.md) for Android
+apps: a `cdylib` exporting `extern "C"` functions (`edgehdf5_*`, a name
+kept from the project's earlier "edgehdf5" days) that manage an
+`HDF5Memory` through an opaque handle.
+
+The functions are plain C symbols, not JNI-mangled `Java_...` entry points:
+a Kotlin/Java app calls them through a thin JNI shim or JNA of its own. No
+such shim, Gradle project or AAR is in this repository, and the crate is
+not built for an Android target in CI (only its host-side unit tests run
+with the workspace).
+
+## Functions
+
+| Function | What |
+|---|---|
+| `edgehdf5_create(path, agent_id, embedding_dim)` / `edgehdf5_open(path)` | a handle, or null on failure |
+| `edgehdf5_close(handle)` | drop the store; what is not yet checkpointed stays in its WAL, as with any `HDF5Memory` |
+| `edgehdf5_save(handle, ...)` | save one entry; the embedding length is checked against the store's dimension before the pointer is read |
+| `edgehdf5_delete`, `edgehdf5_count`, `edgehdf5_count_active` | |
+| `edgehdf5_hybrid_search(handle, query, len, text, vector_weight, keyword_weight, max_results, out_indices, out_scores, out_chunks)` | results into caller-provided arrays; returns the number written |
+| `edgehdf5_add_session`, `edgehdf5_get_session_summary` | sessions |
+| `edgehdf5_add_entity`, `edgehdf5_add_relation` | knowledge graph |
+| `edgehdf5_free_string` | free a string this library returned |
+
+Every function is `unsafe`: the caller guarantees valid, NUL-terminated
+strings and correctly sized buffers (see each function's `# Safety`
+section), and serialises access to a handle; separate handles are
+independent.
+
+## Build
+
+```bash
+cargo build --release -p clawhdf5-android   # host build; for a device, add --target aarch64-linux-android with the NDK's linker configured
+```
+
+It depends on `clawhdf5-agent` with **default features off**, so there is
+no HNSW index (the vector stage is an exact linear scan) and no rayon
+pool. No C is compiled.
+
+## License
+
+MIT

crates/clawhdf5-ann/Cargo.toml

@@ -1,7 +1,8 @@
 [package]
 name = "clawhdf5-ann"
-version = "2.5.0"
+version = "2.7.0"
 edition = "2024"
+rust-version.workspace = true
 description = "HNSW approximate nearest neighbor index stored as HDF5"
 license = "MIT"
 repository = "https://git.redclaw.dev/quantumclaw/clawhdf5"
@@ -10,9 +11,9 @@ keywords = ["hdf5", "ann", "hnsw", "nearest-neighbor"]
 categories = ["algorithms", "science"]
 
 [dependencies]
-clawhdf5-format = { path = "../clawhdf5-format", version = "2.5.0" }
-clawhdf5-io = { path = "../clawhdf5-io", version = "2.5.0" }
-clawhdf5-accel = { path = "../clawhdf5-accel", version = "2.5.0" }
+clawhdf5-format = { path = "../clawhdf5-format", version = "2.7.0" }
+clawhdf5-io = { path = "../clawhdf5-io", version = "2.7.0" }
+clawhdf5-accel = { path = "../clawhdf5-accel", version = "2.7.0" }
 rayon = { version = "1", optional = true }
 
 [features]

crates/clawhdf5-ann/README.md

@@ -1,25 +1,70 @@
 # clawhdf5-ann
 
-[![crates.io](https://img.shields.io/crates/v/clawhdf5-ann.svg)](https://crates.io/crates/clawhdf5-ann)
-[![docs.rs](https://docs.rs/clawhdf5-ann/badge.svg)](https://docs.rs/clawhdf5-ann)
+An HNSW (Hierarchical Navigable Small World) approximate nearest-neighbour
+index in pure Rust, with cosine or L2 distance, optional int8 storage of
+the vectors, deletions, and persistence as an HDF5 file. It is the vector
+stage of [`clawhdf5-agent`](../clawhdf5-agent/README.md)'s search (the
+agent's `hnsw` feature, on by default); distances run on
+[`clawhdf5-accel`](../clawhdf5-accel/README.md)'s SIMD kernels.
 
-HNSW approximate nearest neighbor index stored as HDF5.
+Neighbours are chosen with the HNSW paper's diversity heuristic, not plain
+closest-M (which capped recall on clustered data at 0.31 recall@10 at 100K
+vectors).
 
-## Features
+Not on crates.io yet; depend on it from git:
 
-- Build and query HNSW indexes persisted in HDF5 format
-- Pure Rust, no C dependencies
-- Efficient similarity search for high-dimensional vectors
+```toml
+[dependencies]
+clawhdf5-ann = { git = "https://git.redclaw.dev/quantumclaw/clawhdf5" }
+```
 
 ## Usage
 
 ```rust
-use clawhdf5_ann::HnswIndex;
+use clawhdf5_ann::{DistanceMetric, HnswIndex, Storage};
 
-let index = HnswIndex::from_hdf5("vectors.h5").unwrap();
-let neighbors = index.search(&query, 10);
+let vectors: Vec<Vec<f32>> = (0..500)
+    .map(|i| (0..16).map(|j| ((i * 31 + j * 7) % 97) as f32 / 97.0).collect())
+    .collect();
+
+// m = 16 connections per node, ef_construction = 200
+let mut index = HnswIndex::build_with(&vectors, 16, 200, DistanceMetric::Cosine, Storage::Int8);
+let hits = index.search(&vectors[42], 10, 64); // (id, distance), closest first; ef >= k
+assert!(hits[0].1 < 1e-3); // vector 42 itself (or an identical one)
+
+let id = index.insert(vec![0.5; 16]);
+index.mark_deleted(id);
+
+// Persist as HDF5 (a self-contained file: graph and vectors) and load it back
+let bytes = index.to_hdf5_bytes().unwrap();
+let loaded = HnswIndex::load_from_hdf5(&bytes).unwrap();
+assert_eq!(loaded.len(), index.len());
 ```
 
+- `HnswIndex::build` (L2), `build_with_metric`, `build_with` (metric and
+  storage); `new`/`new_with` plus `insert` for an index built
+  incrementally.
+- `Storage::Int8` keeps each vector as `i8`, a quarter of the memory; it
+  applies to `Cosine` only (an L2 index keeps `Float32`). Distances are then
+  approximate, so a caller that needs exact ranking re-scores the
+  candidates, as the agent does.
+- `mark_deleted`, `is_deleted`, `deleted_count`, `active_len`, `compact`
+  (returns the old-to-new id map).
+- `save_to_hdf5(&mut writer)` / `to_hdf5_bytes` / `load_from_hdf5` store
+  the whole index; `graph_to_bytes` / `from_graph_bytes` store only the
+  graph (with a CRC32) for a caller that keeps the vectors elsewhere — the
+  agent's `<store>.h5.ann` sidecar.
+
+## Features
+
+| Feature | Default | What | Builds C |
+|---|---|---|---|
+| `parallel` | no | build the graph on a rayon pool; the graph is identical with or without it | no |
+
+Recall and speed against exact search, for the index alone and in the
+agent: [`BENCHMARKS.md`](../../BENCHMARKS.md), measured with
+`cargo run --release -p clawhdf5-bench --bin search_harness`.
+
 ## License
 
 MIT

crates/clawhdf5-ann/src/hnsw.rs

@@ -13,7 +13,7 @@ use clawhdf5_format::filter_pipeline::FilterPipeline;
 use clawhdf5_format::group_v2::resolve_path_any;
 use clawhdf5_format::message_type::MessageType;
 use clawhdf5_format::object_header::ObjectHeader;
-use clawhdf5_format::signature::find_signature;
+use clawhdf5_format::signature::split_user_block;
 use clawhdf5_format::superblock::Superblock;
 use clawhdf5_io::FileWriter as IoFileWriter;
 
@@ -154,6 +154,218 @@ impl Ord for FarCandidate {
     }
 }
 
+/// How the index keeps its copy of the vectors.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
+pub enum Storage {
+    /// Exactly as given: `dim * 4` bytes per vector.
+    #[default]
+    Float32,
+    /// Each component scaled to an `i8`: `dim` bytes per vector, a quarter of
+    /// the space, at some cost in precision.
+    ///
+    /// Only meaningful for [`DistanceMetric::Cosine`]: rows are stored
+    /// unit-length, so a quantised dot product reconstructs the similarity
+    /// directly. Requesting it for `L2` keeps `Float32`, because an L2
+    /// distance cannot be recovered from a dot product alone.
+    Int8,
+}
+
+/// The index's copy of the vectors, flat and row-major.
+#[derive(Debug, Clone)]
+enum Vectors {
+    F32 {
+        dim: usize,
+        flat: Vec<f32>,
+    },
+    /// `flat[i * dim + j]` is component `j` of vector `i` divided by
+    /// `scales[i]`; multiplying back recovers it.
+    ///
+    /// The scale is per row rather than global. A unit-length row in `d`
+    /// dimensions has components around `1/sqrt(d)`, so a fixed `[-1, 1]`
+    /// scale spends fewer than 12 of the 255 levels on a 128-dimensional
+    /// vector and the reconstruction error swamps the gaps between near
+    /// neighbours — measured at 0.35 top-10 overlap with the exact ranking.
+    /// Scaling each row by its own largest component uses the full range.
+    Int8 {
+        dim: usize,
+        flat: Vec<i8>,
+        scales: Vec<f32>,
+    },
+}
+
+/// Levels either side of zero. 127, not 128, so the range is symmetric.
+const INT8_LEVELS: f32 = 127.0;
+
+/// Quantise one row, returning the codes and the scale that inverts them.
+fn quantise_row(v: &[f32], out: &mut Vec<i8>) -> f32 {
+    let max_abs = v.iter().fold(0.0f32, |m, x| m.max(x.abs()));
+    if max_abs <= f32::MIN_POSITIVE {
+        out.extend(core::iter::repeat_n(0i8, v.len()));
+        return 0.0;
+    }
+    let inv = INT8_LEVELS / max_abs;
+    out.extend(
+        v.iter()
+            .map(|x| (x * inv).round().clamp(-INT8_LEVELS, INT8_LEVELS) as i8),
+    );
+    max_abs / INT8_LEVELS
+}
+
+impl Vectors {
+    fn new(dim: usize, storage: Storage, metric: DistanceMetric) -> Self {
+        match storage {
+            Storage::Int8 if metric == DistanceMetric::Cosine => Vectors::Int8 {
+                dim,
+                flat: Vec::new(),
+                scales: Vec::new(),
+            },
+            _ => Vectors::F32 {
+                dim,
+                flat: Vec::new(),
+            },
+        }
+    }
+
+    fn dim(&self) -> usize {
+        match self {
+            Vectors::F32 { dim, .. } | Vectors::Int8 { dim, .. } => *dim,
+        }
+    }
+
+    fn storage(&self) -> Storage {
+        match self {
+            Vectors::F32 { .. } => Storage::Float32,
+            Vectors::Int8 { .. } => Storage::Int8,
+        }
+    }
+
+    fn len(&self) -> usize {
+        let dim = self.dim();
+        if dim == 0 {
+            return 0;
+        }
+        match self {
+            Vectors::F32 { flat, .. } => flat.len() / dim,
+            Vectors::Int8 { flat, .. } => flat.len() / dim,
+        }
+    }
+
+    /// Set the row width, for a store seeded empty by `new`.
+    fn set_dim(&mut self, new_dim: usize) {
+        match self {
+            Vectors::F32 { dim, .. } | Vectors::Int8 { dim, .. } => *dim = new_dim,
+        }
+    }
+
+    fn push(&mut self, vector: &[f32]) {
+        match self {
+            Vectors::F32 { flat, .. } => flat.extend_from_slice(vector),
+            Vectors::Int8 { flat, scales, .. } => scales.push(quantise_row(vector, flat)),
+        }
+    }
+
+    /// Row `i` as `f32`, for callers that need the values back (serialization,
+    /// and the f32 fast paths). Quantised rows are reconstructed, so this is
+    /// lossy in exactly the way the storage is.
+    fn row(&self, i: usize) -> Vec<f32> {
+        let dim = self.dim();
+        let start = i * dim;
+        match self {
+            Vectors::F32 { flat, .. } => flat[start..start + dim].to_vec(),
+            Vectors::Int8 { flat, scales, .. } => flat[start..start + dim]
+                .iter()
+                .map(|&q| f32::from(q) * scales[i])
+                .collect(),
+        }
+    }
+
+    /// Distance between two stored vectors.
+    fn dist(&self, a: usize, b: usize, metric: DistanceMetric) -> f32 {
+        let dim = self.dim();
+        match self {
+            Vectors::F32 { flat, .. } => {
+                let (x, y) = (a * dim, b * dim);
+                compute_distance(&flat[x..x + dim], &flat[y..y + dim], metric)
+            }
+            Vectors::Int8 { flat, scales, .. } => {
+                let (x, y) = (a * dim, b * dim);
+                let dot = dot_i8(&flat[x..x + dim], &flat[y..y + dim]);
+                1.0 - dot as f32 * scales[a] * scales[b]
+            }
+        }
+    }
+
+    /// Distance from a prepared query to stored vector `i`.
+    fn dist_query(&self, query: &Query, i: usize, metric: DistanceMetric) -> f32 {
+        let dim = self.dim();
+        let start = i * dim;
+        match (self, query) {
+            (Vectors::F32 { flat, .. }, Query::F32(q)) => {
+                compute_distance(q, &flat[start..start + dim], metric)
+            }
+            (Vectors::Int8 { flat, scales, .. }, Query::Int8(q, q_scale)) => {
+                let dot = dot_i8(q, &flat[start..start + dim]);
+                1.0 - dot as f32 * q_scale * scales[i]
+            }
+            // Mixed forms cannot occur: `Query` is built from the same storage.
+            _ => f32::MAX,
+        }
+    }
+
+    /// Build a store from prepared rows.
+    fn from_rows(rows: &[Vec<f32>], storage: Storage, metric: DistanceMetric) -> Self {
+        let dim = rows.first().map_or(0, Vec::len);
+        let mut out = Vectors::new(dim, storage, metric);
+        for row in rows {
+            out.push(row);
+        }
+        out
+    }
+
+    /// Prepare `query` for comparison against this store.
+    fn query(&self, query: Vec<f32>) -> Query {
+        match self {
+            Vectors::F32 { .. } => Query::F32(query),
+            Vectors::Int8 { .. } => {
+                let mut codes = Vec::with_capacity(query.len());
+                let scale = quantise_row(&query, &mut codes);
+                Query::Int8(codes, scale)
+            }
+        }
+    }
+}
+
+/// What a layer search is measuring distance *to*: an incoming query, or a
+/// node already in the index (which is what insertion compares against).
+enum Target<'a> {
+    Query(&'a Query),
+    Node(usize),
+}
+
+impl Vectors {
+    fn dist_to(&self, target: &Target<'_>, i: usize, metric: DistanceMetric) -> f32 {
+        match target {
+            Target::Query(q) => self.dist_query(q, i, metric),
+            Target::Node(n) => self.dist(*n, i, metric),
+        }
+    }
+}
+
+/// A search query in whichever form the store compares against.
+enum Query {
+    F32(Vec<f32>),
+    /// Codes and the scale that inverts them, as in [`Vectors::Int8`].
+    Int8(Vec<i8>, f32),
+}
+
+/// Sum of products, widened so it cannot overflow. Runtime-dispatched to the
+/// same SIMD backend as the f32 kernels, so the two storages are compared on
+/// equal terms.
+#[inline]
+fn dot_i8(a: &[i8], b: &[i8]) -> i32 {
+    clawhdf5_accel::dot_i8(a, b)
+}
+
 /// Magic for [`HnswIndex::graph_to_bytes`].
 const GRAPH_MAGIC: &[u8; 4] = b"CHG1";
 
@@ -174,8 +386,8 @@ pub const HNSW_FORMAT_VERSION: i64 = 2;
 /// HDF5 format.
 #[derive(Debug, Clone)]
 pub struct HnswIndex {
-    /// All vectors in the index.
-    vectors: Vec<Vec<f32>>,
+    /// All vectors in the index, flat and row-major.
+    vectors: Vectors,
     /// Adjacency lists per layer. `graph[layer][node]` = list of neighbor IDs.
     graph: Vec<Vec<Vec<usize>>>,
     /// Soft-deletion flags, one per node. Deleted nodes remain in the graph for
@@ -214,6 +426,20 @@ impl HnswIndex {
         m: usize,
         ef_construction: usize,
         metric: DistanceMetric,
+    ) -> Self {
+        Self::build_with(vectors, m, ef_construction, metric, Storage::default())
+    }
+
+    /// Build an index, choosing how the vectors are stored.
+    ///
+    /// [`Storage::Int8`] keeps them at a quarter of the size; see its docs for
+    /// what that costs and when it applies.
+    pub fn build_with(
+        vectors: &[Vec<f32>],
+        m: usize,
+        ef_construction: usize,
+        metric: DistanceMetric,
+        storage: Storage,
     ) -> Self {
         assert!(!vectors.is_empty(), "cannot build index from empty vectors");
         assert!(m >= 2, "m must be at least 2");
@@ -224,8 +450,11 @@ impl HnswIndex {
 
         let m_max0 = m * 2;
         let n = vectors.len();
-        let prepared: Vec<Vec<f32>> = vectors.iter().map(|v| prepare(v.clone(), metric)).collect();
-        let vectors: &[Vec<f32>] = &prepared;
+        let mut prepared = Vectors::new(dim, storage, metric);
+        for v in vectors {
+            prepared.push(&prepare(v.clone(), metric));
+        }
+        let vectors = &prepared;
 
         // Assign levels to all nodes
         let mut node_levels = Vec::with_capacity(n);
@@ -322,9 +551,20 @@ impl HnswIndex {
     /// point for incremental [`HnswIndex::insert`] and as the result of
     /// [`HnswIndex::compact`] when every vector has been deleted.
     pub fn new(m: usize, ef_construction: usize, metric: DistanceMetric) -> Self {
+        Self::new_with(m, ef_construction, metric, Storage::default())
+    }
+
+    /// [`HnswIndex::new`], choosing how the vectors are stored.
+    pub fn new_with(
+        m: usize,
+        ef_construction: usize,
+        metric: DistanceMetric,
+        storage: Storage,
+    ) -> Self {
         assert!(m >= 2, "m must be at least 2");
         Self {
-            vectors: Vec::new(),
+            // The dimension is set by the first insert.
+            vectors: Vectors::new(0, storage, metric),
             graph: Vec::new(),
             deleted: Vec::new(),
             entry_point: 0,
@@ -351,7 +591,8 @@ impl HnswIndex {
         // Seed an empty index.
         if id == 0 {
             let node_level = assign_level(0, self.m);
-            self.vectors.push(vector);
+            self.vectors.set_dim(vector.len());
+            self.vectors.push(&vector);
             self.deleted.push(false);
             self.node_levels.push(node_level);
             self.graph = (0..=node_level).map(|_| vec![Vec::new(); 1]).collect();
@@ -361,12 +602,12 @@ impl HnswIndex {
 
         assert_eq!(
             vector.len(),
-            self.vectors[0].len(),
+            self.vectors.dim(),
             "insert dimension mismatch"
         );
 
         let node_level = assign_level(id, self.m);
-        self.vectors.push(vector);
+        self.vectors.push(&vector);
         self.deleted.push(false);
         self.node_levels.push(node_level);
 
@@ -387,7 +628,7 @@ impl HnswIndex {
             ep = greedy_closest(
                 &self.vectors,
                 &self.graph[layer],
-                &self.vectors[id],
+                &Target::Node(id),
                 ep,
                 self.metric,
             );
@@ -400,7 +641,7 @@ impl HnswIndex {
             let neighbors = search_layer(
                 &self.vectors,
                 &self.graph[layer],
-                &self.vectors[id],
+                &Target::Node(id),
                 ep,
                 self.ef_construction,
                 self.metric,
@@ -466,16 +707,25 @@ impl HnswIndex {
     pub fn compact(&mut self) -> Vec<Option<usize>> {
         let mut mapping = vec![None; self.vectors.len()];
         let mut surviving: Vec<Vec<f32>> = Vec::with_capacity(self.active_len());
-        for (old, v) in self.vectors.iter().enumerate() {
+        for (old, slot) in mapping.iter_mut().enumerate() {
             if !self.deleted[old] {
-                mapping[old] = Some(surviving.len());
-                surviving.push(v.clone());
+                *slot = Some(surviving.len());
+                surviving.push(self.vectors.row(old));
             }
         }
+        // Rebuilding must keep the storage the caller chose; a compaction is
+        // not the place to silently quadruple the index's memory.
+        let storage = self.vectors.storage();
         *self = if surviving.is_empty() {
-            Self::new(self.m, self.ef_construction, self.metric)
+            Self::new_with(self.m, self.ef_construction, self.metric, storage)
         } else {
-            Self::build_with_metric(&surviving, self.m, self.ef_construction, self.metric)
+            Self::build_with(
+                &surviving,
+                self.m,
+                self.ef_construction,
+                self.metric,
+                storage,
+            )
         };
         mapping
     }
@@ -490,24 +740,22 @@ impl HnswIndex {
     /// # Returns
     /// A vector of `(id, distance)` pairs sorted by distance (closest first).
     pub fn search(&self, query: &[f32], k: usize, ef: usize) -> Vec<(usize, f32)> {
-        if self.vectors.is_empty() {
+        if self.vectors.len() == 0 {
             return Vec::new();
         }
-        assert_eq!(
-            query.len(),
-            self.vectors[0].len(),
-            "query dimension mismatch"
-        );
+        assert_eq!(query.len(), self.vectors.dim(), "query dimension mismatch");
         let ef = ef.max(k);
-        let prepared_query = prepare(query.to_vec(), self.metric);
-        let query = prepared_query.as_slice();
+        // Prepared and, for a quantised store, quantised once per search
+        // rather than once per comparison.
+        let prepared = self.vectors.query(prepare(query.to_vec(), self.metric));
+        let target = Target::Query(&prepared);
 
         let mut ep = self.entry_point;
         let top_layer = self.graph.len().saturating_sub(1);
 
         // Greedy search from top layer down to layer 1
         for layer in (1..=top_layer).rev() {
-            ep = greedy_closest(&self.vectors, &self.graph[layer], query, ep, self.metric);
+            ep = greedy_closest(&self.vectors, &self.graph[layer], &target, ep, self.metric);
         }
 
         // Search layer 0 for the ef nearest *live* nodes. Deleted nodes are
@@ -516,7 +764,7 @@ impl HnswIndex {
         let candidates = search_layer(
             &self.vectors,
             &self.graph[0],
-            query,
+            &target,
             ep,
             ef,
             self.metric,
@@ -543,14 +791,13 @@ impl HnswIndex {
     pub fn to_hdf5_bytes(&self) -> Result<Vec<u8>, FormatError> {
         let mut fw = FmtWriter::new();
         let n = self.vectors.len();
-        let dim = if n > 0 { self.vectors[0].len() } else { 0 };
+        let dim = self.vectors.dim();
 
         // Flatten vectors into a 1D array for storage
-        let flat_vectors: Vec<f32> = self
-            .vectors
-            .iter()
-            .flat_map(|v| v.iter().copied())
-            .collect();
+        let mut flat_vectors: Vec<f32> = Vec::with_capacity(n * dim);
+        for i in 0..n {
+            flat_vectors.extend_from_slice(&self.vectors.row(i));
+        }
 
         let mut group = fw.create_group("ann");
 
@@ -614,8 +861,9 @@ impl HnswIndex {
     /// The HDF5 data must contain the `/ann/vectors`, `/ann/graph_layer_*`,
     /// and `/ann/config` datasets as produced by [`to_hdf5_bytes`].
     pub fn load_from_hdf5(data: &[u8]) -> Result<Self, FormatError> {
-        let sig_offset = find_signature(data)?;
-        let sb = Superblock::parse(data, sig_offset)?;
+        // Addresses are relative to the superblock: skip any user block.
+        let (_, data) = split_user_block(data)?;
+        let sb = Superblock::parse(data, 0)?;
 
         // Read config dataset and its attributes
         let config_attrs = read_dataset_attrs(data, &sb, "ann/config")?;
@@ -709,7 +957,9 @@ impl HnswIndex {
         };
 
         Ok(Self {
-            vectors,
+            // Serialized files carry f32 vectors and no storage tag: a
+            // quantised index is rebuilt, not loaded.
+            vectors: Vectors::from_rows(&vectors, Storage::Float32, metric),
             graph,
             deleted,
             entry_point,
@@ -773,6 +1023,16 @@ impl HnswIndex {
     /// `bytes` is validated — a corrupt or mismatched graph is an error, never
     /// an index that panics or walks out of bounds during a search.
     pub fn from_graph_bytes(bytes: &[u8], vectors: Vec<Vec<f32>>) -> Result<Self, FormatError> {
+        Self::from_graph_bytes_with(bytes, vectors, Storage::default())
+    }
+
+    /// As [`from_graph_bytes`](Self::from_graph_bytes), choosing how the
+    /// rehydrated vectors are stored.
+    pub fn from_graph_bytes_with(
+        bytes: &[u8],
+        vectors: Vec<Vec<f32>>,
+        storage: Storage,
+    ) -> Result<Self, FormatError> {
         let bad = |what: &str| FormatError::SerializationError(format!("HNSW graph: {what}"));
         let body_len = bytes
             .len()
@@ -863,7 +1123,14 @@ impl HnswIndex {
         }
 
         Ok(Self {
-            vectors: vectors.into_iter().map(|v| prepare(v, metric)).collect(),
+            vectors: Vectors::from_rows(
+                &vectors
+                    .into_iter()
+                    .map(|v| prepare(v, metric))
+                    .collect::<Vec<_>>(),
+                storage,
+                metric,
+            ),
             graph,
             deleted,
             entry_point,
@@ -882,16 +1149,17 @@ impl HnswIndex {
 
     /// Returns true if the index is empty.
     pub fn is_empty(&self) -> bool {
-        self.vectors.is_empty()
+        self.vectors.len() == 0
+    }
+
+    /// How this index stores its copy of the vectors.
+    pub fn storage(&self) -> Storage {
+        self.vectors.storage()
     }
 
     /// Returns the dimension of vectors in the index.
     pub fn dimension(&self) -> usize {
-        if self.vectors.is_empty() {
-            0
-        } else {
-            self.vectors[0].len()
-        }
+        self.vectors.dim()
     }
 
     /// Returns the number of layers in the graph.
@@ -916,17 +1184,17 @@ impl HnswIndex {
 
 /// Greedy search: find the single closest node to `query` starting from `ep`.
 fn greedy_closest(
-    vectors: &[Vec<f32>],
+    vectors: &Vectors,
     layer: &[Vec<usize>],
-    query: &[f32],
+    target: &Target<'_>,
     mut ep: usize,
     metric: DistanceMetric,
 ) -> usize {
-    let mut best_dist = compute_distance(query, &vectors[ep], metric);
+    let mut best_dist = vectors.dist_to(target, ep, metric);
     loop {
         let mut changed = false;
         for &neighbor in &layer[ep] {
-            let d = compute_distance(query, &vectors[neighbor], metric);
+            let d = vectors.dist_to(target, neighbor, metric);
             if d < best_dist {
                 best_dist = d;
                 ep = neighbor;
@@ -950,15 +1218,15 @@ fn greedy_closest(
 /// instead meant a query whose neighbourhood had been deleted got back fewer
 /// than `k` results, or none, however many live records were nearby.
 fn search_layer(
-    vectors: &[Vec<f32>],
+    vectors: &Vectors,
     layer: &[Vec<usize>],
-    query: &[f32],
+    target: &Target<'_>,
     ep: usize,
     ef: usize,
     metric: DistanceMetric,
     skip: Option<&[bool]>,
 ) -> Vec<Candidate> {
-    let ep_dist = compute_distance(query, &vectors[ep], metric);
+    let ep_dist = vectors.dist_to(target, ep, metric);
 
     // Min-heap of candidates to explore
     let mut candidates = BinaryHeap::new();
@@ -980,7 +1248,7 @@ fn search_layer(
         visited.begin(vectors.len());
         visited.insert(ep);
         search_layer_visit(
-            vectors, layer, query, ef, metric, skip, visited, candidates, results,
+            vectors, layer, target, ef, metric, skip, visited, candidates, results,
         )
     })
 }
@@ -1023,9 +1291,9 @@ thread_local! {
 
 #[allow(clippy::too_many_arguments)]
 fn search_layer_visit(
-    vectors: &[Vec<f32>],
+    vectors: &Vectors,
     layer: &[Vec<usize>],
-    query: &[f32],
+    target: &Target<'_>,
     ef: usize,
     metric: DistanceMetric,
     skip: Option<&[bool]>,
@@ -1044,7 +1312,7 @@ fn search_layer_visit(
                 continue;
             }
 
-            let d = compute_distance(query, &vectors[neighbor], metric);
+            let d = vectors.dist_to(target, neighbor, metric);
             let furthest_dist = results.peek().map_or(f32::MAX, |f| f.distance);
 
             if d < furthest_dist || results.len() < ef {
@@ -1095,7 +1363,7 @@ fn search_layer_visit(
 /// remaining slots are then filled with the closest rejected candidates, so a
 /// node is never left under-connected.
 fn select_neighbors(
-    vectors: &[Vec<f32>],
+    vectors: &Vectors,
     candidates: &[(usize, f32)],
     max_conn: usize,
     metric: DistanceMetric,
@@ -1111,7 +1379,7 @@ fn select_neighbors(
         }
         let diverse = selected
             .iter()
-            .all(|&s| compute_distance(&vectors[id], &vectors[s], metric) > dist_to_node);
+            .all(|&s| vectors.dist(id, s, metric) > dist_to_node);
         if diverse {
             selected.push(id);
         } else {
@@ -1136,7 +1404,7 @@ fn batch_len(linked: usize) -> usize {
 /// layers, found by searching the graph as it currently stands.
 #[allow(clippy::too_many_arguments)]
 fn plan_batch(
-    vectors: &[Vec<f32>],
+    vectors: &Vectors,
     graph: &[Vec<Vec<usize>>],
     node_levels: &[usize],
     batch: std::ops::Range<usize>,
@@ -1150,7 +1418,7 @@ fn plan_batch(
         let mut ep = entry_point;
         // Phase 1: greedy descent from the top layer down to node_level + 1.
         for layer in (node_level + 1..=ep_level).rev() {
-            ep = greedy_closest(vectors, &graph[layer], &vectors[i], ep, metric);
+            ep = greedy_closest(vectors, &graph[layer], &Target::Node(i), ep, metric);
         }
         // Phase 2: search and select on every layer the node lives on.
         let mut plan = Vec::with_capacity(node_level.min(ep_level) + 1);
@@ -1159,7 +1427,7 @@ fn plan_batch(
             let neighbors = search_layer(
                 vectors,
                 &graph[layer],
-                &vectors[i],
+                &Target::Node(i),
                 ep,
                 ef_construction,
                 metric,
@@ -1186,7 +1454,7 @@ fn plan_batch(
 /// Prune every `(layer, node)` neighbour list in `overflowed` back to its
 /// limit. Each list belongs to a different node, so they are independent.
 fn prune_overflowed(
-    vectors: &[Vec<f32>],
+    vectors: &Vectors,
     graph: &mut [Vec<Vec<usize>>],
     overflowed: Vec<(usize, usize)>,
     (m, m_max0): (usize, usize),
@@ -1226,7 +1494,7 @@ const PARALLEL_MIN: usize = 8;
 /// prunes is too fine-grained to parallelise profitably — measured 1.45x on 16
 /// cores; bulk builds batch their pruning instead, see `prune_overflowed`.)
 fn link_back(
-    vectors: &[Vec<f32>],
+    vectors: &Vectors,
     layer: &mut [Vec<usize>],
     new_id: usize,
     selected: &[usize],
@@ -1248,7 +1516,7 @@ fn link_back(
 
 /// Trim `node`'s neighbour list back to `max_conn` with [`select_neighbors`].
 fn prune_connections(
-    vectors: &[Vec<f32>],
+    vectors: &Vectors,
     neighbors: &mut Vec<usize>,
     node: usize,
     max_conn: usize,
@@ -1259,7 +1527,7 @@ fn prune_connections(
     }
     let mut scored: Vec<(usize, f32)> = neighbors
         .iter()
-        .map(|&n| (n, compute_distance(&vectors[node], &vectors[n], metric)))
+        .map(|&n| (n, vectors.dist(node, n, metric)))
         .collect();
     scored.sort_by(|a, b| a.1.total_cmp(&b.1).then(a.0.cmp(&b.0)));
     *neighbors = select_neighbors(vectors, &scored, max_conn, metric);
@@ -1519,6 +1787,88 @@ mod tests {
         assert!(recall >= 0.95, "incremental recall@10 = {recall}");
     }
 
+    #[test]
+    fn int8_storage_needs_an_exact_re_score_to_match_f32() {
+        // Cosine only: rows are unit-length, so a quantised dot product
+        // reconstructs the similarity directly.
+        //
+        // Not the `clustered` generator: its clusters are far tighter than any
+        // real embedding, so neighbours sit closer together than the
+        // quantisation error and top-10 identity there is noise — that would
+        // measure the fixture, not the storage.
+        let mut vectors = make_random_vectors(3060, 128, 5);
+        let queries = vectors.split_off(3000);
+        let f32_index =
+            HnswIndex::build_with(&vectors, 8, 40, DistanceMetric::Cosine, Storage::Float32);
+        let quantised =
+            HnswIndex::build_with(&vectors, 8, 40, DistanceMetric::Cosine, Storage::Int8);
+        assert_eq!(quantised.storage(), Storage::Int8);
+
+        // Ground truth, not the f32 index's answers: re-scoring can beat that
+        // index, and measuring against it would score being right as drift.
+        let truth: Vec<Vec<usize>> = queries
+            .iter()
+            .map(|q| {
+                let mut d: Vec<(usize, f32)> = vectors
+                    .iter()
+                    .enumerate()
+                    .map(|(i, v)| (i, compute_distance(q, v, DistanceMetric::Cosine)))
+                    .collect();
+                d.sort_by(|a, b| a.1.total_cmp(&b.1));
+                d[..10].iter().map(|x| x.0).collect()
+            })
+            .collect();
+
+        let recall = |got: &dyn Fn(&[f32]) -> Vec<usize>| -> f64 {
+            let mut hits = 0;
+            for (q, want) in queries.iter().zip(&truth) {
+                hits += got(q).iter().filter(|id| want.contains(id)).count();
+            }
+            hits as f64 / (10 * queries.len()) as f64
+        };
+
+        let exact_recall = recall(&|q| f32_index.search(q, 10, 64).iter().map(|r| r.0).collect());
+        let raw_recall = recall(&|q| quantised.search(q, 10, 64).iter().map(|r| r.0).collect());
+        // Quantised distances alone cost recall, and `ef` cannot buy it back:
+        // the loss is in the distances, not in the graph.
+        assert!(
+            raw_recall < exact_recall,
+            "int8 alone should cost recall: {raw_recall} vs {exact_recall}"
+        );
+
+        // Re-scoring a wider candidate pool against the exact vectors — what a
+        // caller holding them (the agent's embedding cache) does — puts it
+        // back, because only the *ordering* was approximate.
+        let rescored_recall = recall(&|q| {
+            let mut pool: Vec<(usize, f32)> = quantised
+                .search(q, 40, 64)
+                .into_iter()
+                .map(|(id, _)| {
+                    (
+                        id,
+                        compute_distance(q, &vectors[id], DistanceMetric::Cosine),
+                    )
+                })
+                .collect();
+            pool.sort_by(|a, b| a.1.total_cmp(&b.1));
+            pool.truncate(10);
+            pool.into_iter().map(|p| p.0).collect()
+        });
+        assert!(
+            rescored_recall >= exact_recall - 0.01,
+            "int8 + exact re-score should match f32: {rescored_recall} vs {exact_recall} (raw {raw_recall})"
+        );
+    }
+
+    #[test]
+    fn int8_storage_falls_back_to_f32_for_non_cosine_metrics() {
+        // L2 distance is not recoverable from a quantised dot product, so the
+        // store silently stays f32 rather than returning wrong distances.
+        let vectors = clustered(100, 8, 5, 3);
+        let index = HnswIndex::build_with(&vectors, 8, 40, DistanceMetric::L2, Storage::Int8);
+        assert_eq!(index.storage(), Storage::Float32);
+    }
+
     #[test]
     fn deletions_near_the_query_do_not_shrink_or_degrade_results() {
         let mut vectors = clustered(2040, 16, 20, 11);
@@ -1650,6 +2000,7 @@ mod tests {
             vec![1.2, 0.0],  // 3
             vec![-2.0, 0.0], // 4
         ];
+        let store = Vectors::from_rows(&vectors, Storage::Float32, DistanceMetric::L2);
         let scored: Vec<(usize, f32)> = (1..5)
             .map(|i| {
                 (
@@ -1659,12 +2010,12 @@ mod tests {
             })
             .collect();
         assert_eq!(
-            select_neighbors(&vectors, &scored, 2, DistanceMetric::L2),
+            select_neighbors(&store, &scored, 2, DistanceMetric::L2),
             [1, 4]
         );
         // Spare capacity is filled with the closest rejected candidates.
         assert_eq!(
-            select_neighbors(&vectors, &scored, 3, DistanceMetric::L2),
+            select_neighbors(&store, &scored, 3, DistanceMetric::L2),
             [1, 4, 2]
         );
     }
@@ -1790,7 +2141,7 @@ mod tests {
 
         // Verify vectors match
         for i in 0..loaded.len() {
-            assert_eq!(loaded.vectors[i], index.vectors[i]);
+            assert_eq!(loaded.vectors.row(i), index.vectors.row(i));
         }
     }
 

crates/clawhdf5-ann/src/lib.rs

@@ -5,4 +5,4 @@
 
 mod hnsw;
 
-pub use hnsw::{DistanceMetric, HnswIndex};
+pub use hnsw::{DistanceMetric, HnswIndex, Storage};

crates/clawhdf5-bench/Cargo.toml

@@ -1,7 +1,8 @@
 [package]
 name = "clawhdf5-bench"
-version = "2.5.0"
+version = "2.7.0"
 edition = "2024"
+rust-version.workspace = true
 description = "Benchmark harnesses for clawhdf5-agent (Track 8)"
 license = "MIT"
 
@@ -33,6 +34,10 @@ path = "src/bin/consolidation_efficiency.rs"
 name = "ephemeral_perf"
 path = "src/bin/ephemeral_perf.rs"
 
+[[bin]]
+name = "concurrent_read"
+path = "src/bin/concurrent_read.rs"
+
 [[bin]]
 name = "mpi_io_bench"
 path = "src/bin/mpi_io_bench.rs"
@@ -63,6 +68,10 @@ clawhdf5-io = { path = "../clawhdf5-io" }
 mpi = { version = "0.8", optional = true }
 serde = { workspace = true }
 serde_json = "1"
+# concurrent_read: size the decode pool (--decode-threads) and evict files
+# from the page cache (--cold, posix_fadvise). Both pure Rust / bindings only.
+rayon = "1"
+libc = "0.2"
 tempfile = { workspace = true }
 # Optional: libhdf5 C wrapper for side-by-side comparison (requires system libhdf5).
 # Enable with: cargo bench -p clawhdf5-bench --features libhdf5-compare

crates/clawhdf5-bench/README.md

@@ -0,0 +1,49 @@
+# clawhdf5-bench
+
+The measurement harnesses behind [`BENCHMARKS.md`](../../BENCHMARKS.md):
+HDF5 read and write speed (against libhdf5 and h5py where noted) and the
+agent store's search, footprint and retrieval quality. Not meant for
+publishing; nothing else in the workspace depends on it. Run everything with
+`--release`, and quote numbers with the machine, date and command, as
+`BENCHMARKS.md` does.
+
+## Binaries
+
+| Binary | Measures |
+|---|---|
+| `read_harness` | full reads vs hyperslab selections of a chunked 2-D dataset (compressed and not) and a contiguous one: does a selection cost scale with the selection or the dataset? (`-- --large` for 512 MB) |
+| `concurrent_read` | decoded read throughput vs threads on one open `File`; `scripts/concurrent_read_h5py.py` runs the same workload with h5py (threads and processes) and `scripts/compare_concurrent_read.py` tabulates both |
+| `search_harness` | HNSW recall@10 vs exact search, QPS and latency per `ef`, and end-to-end `HDF5Memory` ingest/checkpoint/open/search at 1K–100K (`--full`); studies: `--float16-study`, `--options-study`, `--signing-study`, `--ann-only --uniform` |
+| `longmemeval_bench` | LongMemEval retrieval recall (turn and session Hit@k, MRR) — **retrieval, not QA accuracy**. Oracle or full `longmemeval_s` haystack; `--features embeddings` (or `embeddings-cuda`) embeds with MiniLM, otherwise the vector stage is inert and the run is BM25-only |
+| `memory_arena` | a deterministic multi-session retrieval benchmark (BM25-only) |
+| `footprint_bench` | file size and bytes per record at 100–100K records, float16 or `--f32`, WAL on/off, compressed or not |
+| `consolidation_efficiency` | retrieval before and after consolidation on signal + noise records |
+| `ephemeral_perf` | the in-memory ephemeral tier's set/get latency |
+| `mpi_io_bench` | `clawhdf5-io`'s `MpiVol` (root-read + broadcast, not collective I/O); needs `--features mpi-io` and `mpirun` |
+
+```bash
+cargo run --release -p clawhdf5-bench --bin search_harness -- --full
+cargo run --release -p clawhdf5-bench --bin read_harness
+```
+
+## Criterion benches and example
+
+- `cargo bench -p clawhdf5-bench` runs `h5bench_write`, `h5bench_read` and
+  `h5bench_meta` (h5bench-style sequential, chunked, strided and metadata
+  workloads). `--features libhdf5-compare` adds the same workloads through
+  libhdf5 (the `hdf5-metno` crate; needs a system libhdf5 1.14).
+- `examples/worldmodel_sampling.rs`: shuffled per-frame reads of a
+  `(N, H, W, C)` `uint8` dataset, clawhdf5 against h5py on the same file.
+
+## Features
+
+| Feature | What | Builds C |
+|---|---|---|
+| `libhdf5-compare` | libhdf5 variants of the Criterion benches | links the system libhdf5 |
+| `mpi-io` | `mpi_io_bench` | yes (`mpi-sys`; needs an MPI installation) |
+| `embeddings` | MiniLM embeddings for `longmemeval_bench` (candle) | yes (a `cc` build dependency in the candle/tokenizers tree) |
+| `embeddings-cuda` | the same on a CUDA GPU (minutes instead of hours on the full haystack) | yes (CUDA) |
+
+## License
+
+MIT

crates/clawhdf5-bench/scripts/compare_concurrent_read.py

@@ -0,0 +1,70 @@
+#!/usr/bin/env python3
+"""Tabulate concurrent_read JSON results (clawhdf5, h5py threads/processes).
+
+    python compare_concurrent_read.py clawhdf5.json h5py-threads.json h5py-procs.json
+
+Prints one Markdown table: for each layout, mode and thread count, every
+tool's MB/s and scaling efficiency, and the first file's MB/s relative to each
+of the others. Refuses to compare runs whose workload parameters differ.
+"""
+
+import json
+import sys
+
+COMPARED = ("datasets", "rows", "cols", "chunk", "deflate_level", "slab", "slabs", "seed")
+
+
+def main(paths):
+    if len(paths) < 2:
+        sys.exit(__doc__)
+    docs = []
+    for p in paths:
+        with open(p) as fh:
+            docs.append(json.load(fh))
+    ref = docs[0]
+    for d, p in zip(docs[1:], paths[1:]):
+        diff = [k for k in COMPARED if d["params"].get(k) != ref["params"].get(k)]
+        if diff:
+            sys.exit(f"{p}: workload differs from {paths[0]} in {', '.join(diff)}")
+        if d["cache"] != ref["cache"]:
+            print(f"warning: {p} ran {d['cache']!r}, {paths[0]} ran {ref['cache']!r}",
+                  file=sys.stderr)
+        if d.get("host") != ref.get("host"):
+            print(f"warning: {p} ran on {d.get('host')}, {paths[0]} on {ref.get('host')}",
+                  file=sys.stderr)
+
+    names = [d["tool"] for d in docs]
+    for d in docs:
+        extra = f", HDF5 {d['hdf5_version']}" if "hdf5_version" in d else ""
+        print(f"- {d['tool']} {d['version']}{extra}: host {d.get('host')}, "
+              f"{d.get('cpus')} CPUs, cache {d['cache']}, decode threads per read "
+              f"{d.get('decode_threads')}")
+    p = ref["params"]
+    print(f"\n{p['datasets']} datasets of {p['rows']} x {p['cols']} f32, chunks "
+          f"{p['chunk'][0]} x {p['chunk'][1]} (deflate {p['deflate_level']}); "
+          f"`same`: {p['slabs']} slabs of {p['slab']} x {p['slab']}\n")
+
+    index = [{(r["layout"], r["mode"], r["threads"]): r for r in d["results"]} for d in docs]
+    keys = [(r["layout"], r["mode"], r["threads"]) for r in ref["results"]]
+
+    head = ["layout", "mode", "threads"]
+    head += [f"{n} MB/s (eff)" for n in names]
+    head += [f"{names[0]} / {n}" for n in names[1:]]
+    print("| " + " | ".join(head) + " |")
+    print("|---|---|" + "---:|" * (len(head) - 2))
+    for key in keys:
+        cells = [key[0], key[1], str(key[2])]
+        rs = [ix.get(key) for ix in index]
+        for r in rs:
+            if r is None:
+                cells.append("-")
+            else:
+                eff = "-" if r["efficiency"] is None else f"{r['efficiency']:.2f}"
+                cells.append(f"{r['mb_s']:.0f} ({eff})")
+        for r in rs[1:]:
+            cells.append("-" if r is None else f"{rs[0]['mb_s'] / r['mb_s']:.2f}x")
+        print("| " + " | ".join(cells) + " |")
+
+
+if __name__ == "__main__":
+    main(sys.argv[1:])

crates/clawhdf5-bench/scripts/concurrent_read_h5py.py

@@ -0,0 +1,265 @@
+#!/usr/bin/env python3
+"""The concurrent_read workload with h5py, on the files concurrent_read wrote.
+
+libhdf5 serialises every API call under one global lock, and h5py holds its
+own global lock around every call as well, so h5py *threads* cannot decode in
+parallel. h5py users scale with *processes* instead; ``--executor processes``
+measures that (each worker opens the file itself).
+
+The workload mirrors ``crates/clawhdf5-bench/src/bin/concurrent_read.rs``:
+
+* ``distinct``: every dataset read in full once per repetition; worker ``t``
+  of ``T`` reads datasets ``t, t + T, ...``.
+* ``same``: ``--slabs`` random ``--slab`` x ``--slab`` hyperslabs of ``d00``
+  (slab ``j`` to worker ``j % T``), offsets from the same splitmix64 stream.
+
+Each worker times itself from a start barrier; a repetition spans the earliest
+start to the latest finish (CLOCK_MONOTONIC, comparable across processes).
+Threads share one ``h5py.File`` per repetition; process workers open the file
+inside the timed region (a few ms against reads of many MiB).
+
+Generate the files first with the Rust harness (it writes ``manifest.json``),
+then, for example::
+
+    python concurrent_read_h5py.py --dir DIR --executor threads --json h5py-threads.json
+    python concurrent_read_h5py.py --dir DIR --executor processes --json h5py-procs.json
+"""
+
+import argparse
+import json
+import multiprocessing as mp
+import os
+import platform
+import socket
+import sys
+import threading
+import time
+
+import h5py
+import numpy as np
+
+M64 = (1 << 64) - 1
+
+
+def splitmix64(state):
+    """Return (new_state, value); the same stream as the Rust harness."""
+    state = (state + 0x9E3779B97F4A7C15) & M64
+    z = state
+    z = ((z ^ (z >> 30)) * 0xBF58476D1CE4E5B9) & M64
+    z = ((z ^ (z >> 27)) * 0x94D049BB133111EB) & M64
+    return state, z ^ (z >> 31)
+
+
+def value(k, i):
+    """Element i (row-major) of dataset k, exactly as concurrent_read writes it."""
+    _, noise = splitmix64(i ^ (k << 40))
+    return np.float32((((i >> 6) % 16384) + k) + (noise & 0xFF) / 256.0)
+
+
+def slab_offsets(seed, count, rows, cols, slab):
+    s = seed
+    out = []
+    for _ in range(count):
+        s, r = splitmix64(s)
+        s, c = splitmix64(s)
+        out.append((r % (rows - slab + 1), c % (cols - slab + 1)))
+    return out
+
+
+def now():
+    return time.clock_gettime(time.CLOCK_MONOTONIC)
+
+
+def work(f, mode, t, threads, m, slabs, slab, verify):
+    """Worker t's share of one repetition on an open h5py.File."""
+    n = m["rows"] * m["cols"]
+    if mode == "distinct":
+        for k in range(t, m["datasets"], threads):
+            got = f[f"d{k:02d}"][...]
+            assert got.size == n
+            if verify:
+                flat = got.reshape(-1)
+                for i in (0, n // 3, n - 1):
+                    assert flat[i] == value(k, i), f"d{k:02d}[{i}]"
+    else:
+        ds = f["d00"]
+        cols = m["cols"]
+        for r, c in slabs[t::threads]:
+            got = ds[r : r + slab, c : c + slab]
+            assert got.shape == (slab, slab)
+            if verify:
+                assert got[0, 0] == value(0, r * cols + c)
+                last = (r + slab - 1) * cols + c + slab - 1
+                assert got[-1, -1] == value(0, last)
+
+
+# ----- process workers ------------------------------------------------------
+
+_barrier = None
+
+
+def _init(barrier):
+    global _barrier
+    _barrier = barrier
+
+
+def _proc_task(task):
+    path, mode, t, threads, m, slabs, slab = task
+    _barrier.wait()
+    start = now()
+    with h5py.File(path, "r") as f:
+        work(f, mode, t, threads, m, slabs, slab, False)
+    return start, now()
+
+
+def _noop(_):
+    return os.getpid()
+
+
+def run_threads(path, mode, threads, m, slabs, slab):
+    spans = [None] * threads
+    barrier = threading.Barrier(threads)
+    with h5py.File(path, "r") as f:
+
+        def body(t):
+            barrier.wait()
+            start = now()
+            work(f, mode, t, threads, m, slabs, slab, False)
+            spans[t] = (start, now())
+
+        ts = [threading.Thread(target=body, args=(t,)) for t in range(threads)]
+        for th in ts:
+            th.start()
+        for th in ts:
+            th.join()
+    return max(e for _, e in spans) - min(s for s, _ in spans)
+
+
+def run_processes(pool, path, mode, threads, m, slabs, slab):
+    tasks = [(path, mode, t, threads, m, slabs, slab) for t in range(threads)]
+    # One task per worker: each blocks in the barrier until all T have
+    # started, so no worker can take a second task.
+    spans = pool.map(_proc_task, tasks, chunksize=1)
+    return max(e for _, e in spans) - min(s for s, _ in spans)
+
+
+def warm(path):
+    with open(path, "rb") as fh:
+        while fh.read(1 << 24):
+            pass
+
+
+def evict(path):
+    fd = os.open(path, os.O_RDONLY)
+    try:
+        os.posix_fadvise(fd, 0, 0, os.POSIX_FADV_DONTNEED)
+    finally:
+        os.close(fd)
+
+
+def main():
+    ap = argparse.ArgumentParser(description=__doc__.split("\n\n")[0])
+    ap.add_argument("--dir", default="concurrent-read-data")
+    ap.add_argument("--executor", choices=["threads", "processes"], default="threads")
+    ap.add_argument("--threads", default="1,2,4,8,16")
+    ap.add_argument("--reps", type=int, default=3)
+    ap.add_argument("--slab", type=int, default=256)
+    ap.add_argument("--slabs", type=int, default=1024)
+    ap.add_argument("--seed", type=int, default=42)
+    ap.add_argument("--cold", action="store_true")
+    ap.add_argument("--modes", default="distinct,same")
+    ap.add_argument("--layouts", default="deflate,contiguous")
+    ap.add_argument("--json")
+    a = ap.parse_args()
+
+    # The Rust harness pins this value (splitmix64_reference).
+    assert splitmix64(42)[1] == 0xBDD732262FEB6E95, "splitmix64 port is wrong"
+
+    try:
+        with open(os.path.join(a.dir, "manifest.json")) as fh:
+            m = json.load(fh)
+    except FileNotFoundError:
+        sys.exit(f"{a.dir}/manifest.json not found: generate the files with "
+                 "`cargo run --release -p clawhdf5-bench --bin concurrent_read -- --dir ...` first")
+    threads_list = [int(x) for x in a.threads.split(",")]
+    modes = a.modes.split(",")
+    layouts = a.layouts.split(",")
+    if a.slab < 1 or a.slab > min(m["rows"], m["cols"]):
+        sys.exit(f"--slab must be 1..={min(m['rows'], m['cols'])}")
+    files = dict(m["files"])
+    slabs = slab_offsets(a.seed, a.slabs, m["rows"], m["cols"], a.slab)
+    dataset_bytes = m["rows"] * m["cols"] * 4
+    tool = f"h5py-{a.executor}"
+
+    ctx = mp.get_context("spawn")  # never fork a process holding HDF5 state
+    pools = {}
+    if a.executor == "processes":
+        for t in threads_list:
+            pool = ctx.Pool(t, initializer=_init, initargs=(ctx.Barrier(t),))
+            pool.map(_noop, range(t))  # start the workers outside the timing
+            pools[t] = pool
+
+    rows = []
+    print("| layout | mode | threads | MB/s | efficiency | median s |")
+    print("|---|---|---:|---:|---:|---:|")
+    try:
+        for layout in layouts:
+            path = os.path.join(a.dir, files[layout])
+            if not a.cold:
+                warm(path)
+            for mode in modes:
+                with h5py.File(path, "r") as f:  # untimed, checked pass
+                    work(f, mode, 0, 1, m, slabs, a.slab, True)
+                nbytes = (dataset_bytes * m["datasets"] if mode == "distinct"
+                          else a.slab * a.slab * 4 * a.slabs)
+                base = None
+                for t in threads_list:
+                    times = []
+                    for _ in range(a.reps):
+                        if a.cold:
+                            evict(path)
+                        if a.executor == "threads":
+                            times.append(run_threads(path, mode, t, m, slabs, a.slab))
+                        else:
+                            times.append(run_processes(pools[t], path, mode, t, m, slabs, a.slab))
+                    med = sorted(times)[len(times) // 2]
+                    mb_s = nbytes / (1 << 20) / med
+                    if t == 1:
+                        base = mb_s
+                    eff = mb_s / (t * base) if base else None
+                    print(f"| {layout} | {mode} | {t} | {mb_s:.0f} | "
+                          f"{'-' if eff is None else f'{eff:.2f}'} | {med:.4f} |")
+                    rows.append({
+                        "layout": layout, "mode": mode, "threads": t, "bytes": nbytes,
+                        "times_s": times, "median_s": med, "mb_s": mb_s, "efficiency": eff,
+                    })
+    finally:
+        for pool in pools.values():
+            pool.terminate()
+
+    if a.json:
+        doc = {
+            "tool": tool,
+            "version": h5py.__version__,
+            "hdf5_version": h5py.version.hdf5_version,
+            "python": platform.python_version(),
+            "host": socket.gethostname(),
+            "cpus": os.cpu_count(),
+            "unix_time": int(time.time()),
+            "cache": ("cold (posix_fadvise DONTNEED before each repetition)"
+                      if a.cold else "warm"),
+            "decode_threads": 1,
+            "params": {
+                "datasets": m["datasets"], "rows": m["rows"], "cols": m["cols"],
+                "chunk": m["chunk"], "deflate_level": m["deflate_level"],
+                "mib": dataset_bytes // (1 << 20), "slab": a.slab, "slabs": a.slabs,
+                "seed": a.seed, "reps": a.reps, "dir": a.dir,
+            },
+            "results": rows,
+        }
+        with open(a.json, "w") as fh:
+            json.dump(doc, fh, indent=2)
+
+
+if __name__ == "__main__":
+    main()

crates/clawhdf5-bench/src/bin/concurrent_read.rs

@@ -0,0 +1,523 @@
+//! Concurrent-read harness: how does decoded read throughput scale with the
+//! number of threads reading one open file?
+//!
+//! libhdf5 (threadsafe build) serialises every API call under one global
+//! mutex, and h5py holds it too, so threads cannot decode in parallel there.
+//! A clawhdf5 [`File`] is `Send + Sync`; this harness measures what that buys.
+//! `crates/clawhdf5-bench/scripts/concurrent_read_h5py.py` runs the same
+//! workload on the same files with h5py (threads, and processes), and
+//! `compare_concurrent_read.py` tabulates the JSON both write.
+//!
+//! Files (generated on first use, reused while `manifest.json` matches):
+//!
+//! * `<dir>/deflate.h5`: `--datasets` datasets `d00`, `d01`, ... of `f32`,
+//!   `--mib` MiB decoded each, shape `[mib * 256, 1024]`, chunks `256 x 256`,
+//!   deflate level 4.
+//! * `<dir>/contiguous.h5`: the same datasets, contiguous.
+//!
+//! Modes, for each layout and each thread count `T` (strong scaling: the total
+//! work per repetition is fixed, split among the threads):
+//!
+//! * `distinct`: every dataset is read in full once; thread `t` reads datasets
+//!   `t, t + T, t + 2T, ...`.
+//! * `same`: all threads read `d00`, `--slabs` random `--slab` x `--slab`
+//!   hyperslabs in total (slab `j` goes to thread `j % T`). The offsets come
+//!   from a splitmix64 stream seeded with `--seed`, identical in the h5py
+//!   script.
+//!
+//! One `File` per layout per repetition is shared by all threads (opened
+//! fresh each repetition, so no chunk cache carries over). Page cache:
+//! `warm` (default) reads every file once before timing; `--cold` evicts the
+//! files from the page cache with `posix_fadvise(POSIX_FADV_DONTNEED)` before
+//! every repetition (no root needed; it only evicts clean, unmapped pages, so
+//! it is best effort — the JSON says which was used).
+//!
+//! Decode inside one read is itself parallel when clawhdf5-format's `parallel`
+//! feature is on (it is in this binary, via clawhdf5-agent). `--decode-threads
+//! N` sizes that rayon pool; `--decode-threads 1` measures the API's own
+//! thread scaling, comparable with h5py where each call decodes on the
+//! calling thread.
+//!
+//! ```text
+//! cargo run --release -p clawhdf5-bench --bin concurrent_read -- \
+//!     --dir /data/concurrent-read --json clawhdf5.json
+//! cargo run --release -p clawhdf5-bench --bin concurrent_read -- \
+//!     --dir /tmp/cr --datasets 4 --mib 1 --threads 1,2 --slabs 16 --reps 1   # smoke
+//! ```
+
+use std::path::{Path, PathBuf};
+use std::sync::Barrier;
+use std::time::Instant;
+
+use clawhdf5::{File, FileBuilder, Selection};
+use serde::{Deserialize, Serialize};
+
+const COLS: u64 = 1024;
+const ROWS_PER_MIB: u64 = 256; // 256 rows x 1024 cols x 4 bytes = 1 MiB
+const CHUNK: u64 = 256;
+const DEFLATE_LEVEL: u32 = 4;
+const LAYOUTS: [&str; 2] = ["deflate", "contiguous"];
+const MANIFEST_VERSION: u32 = 1;
+
+/// splitmix64 — shared with the h5py script, which must produce the same
+/// stream (both the data and the hyperslab offsets depend on it).
+fn splitmix64(state: &mut u64) -> u64 {
+    *state = state.wrapping_add(0x9E37_79B9_7F4A_7C15);
+    let mut z = *state;
+    z = (z ^ (z >> 30)).wrapping_mul(0xBF58_476D_1CE4_E5B9);
+    z = (z ^ (z >> 27)).wrapping_mul(0x94D0_49BB_1331_11EB);
+    z ^ (z >> 31)
+}
+
+/// Element `i` (row-major) of dataset `k`: a slowly varying integer part plus
+/// 8 bits of noise, so deflate has real work to do (about 3.1x) and every value
+/// is exact in `f32` (< 2^15 with 8 fraction bits), which lets both harnesses
+/// check what they read against this formula.
+fn value(k: u64, i: u64) -> f32 {
+    let mut s = i ^ (k << 40);
+    let noise = splitmix64(&mut s) & 0xff;
+    (((i >> 6) % 16384) + k) as f32 + noise as f32 / 256.0
+}
+
+#[derive(Serialize, Deserialize, PartialEq, Debug, Clone)]
+struct Manifest {
+    version: u32,
+    datasets: u64,
+    rows: u64,
+    cols: u64,
+    chunk: [u64; 2],
+    deflate_level: u32,
+    files: Vec<(String, String)>, // (layout, file name)
+    writer: String,
+}
+
+fn manifest_for(datasets: u64, mib: u64) -> Manifest {
+    Manifest {
+        version: MANIFEST_VERSION,
+        datasets,
+        rows: mib * ROWS_PER_MIB,
+        cols: COLS,
+        chunk: [CHUNK, CHUNK],
+        deflate_level: DEFLATE_LEVEL,
+        files: LAYOUTS
+            .iter()
+            .map(|l| (l.to_string(), format!("{l}.h5")))
+            .collect(),
+        writer: format!("clawhdf5 {}", env!("CARGO_PKG_VERSION")),
+    }
+}
+
+fn dataset_values(k: u64, n: u64) -> Vec<f32> {
+    (0..n).map(|i| value(k, i)).collect()
+}
+
+/// Write the files unless `dir` already holds ones matching `want`.
+fn ensure_files(dir: &Path, want: &Manifest) -> std::io::Result<bool> {
+    let manifest_path = dir.join("manifest.json");
+    if let Ok(text) = std::fs::read_to_string(&manifest_path)
+        && let Ok(have) = serde_json::from_str::<Manifest>(&text)
+        && have.version == want.version
+        && have.datasets == want.datasets
+        && have.rows == want.rows
+        && have.cols == want.cols
+        && have.chunk == want.chunk
+        && have.deflate_level == want.deflate_level
+        && have.files == want.files
+        && want.files.iter().all(|(_, f)| dir.join(f).exists())
+    {
+        return Ok(false);
+    }
+    std::fs::create_dir_all(dir)?;
+    // A stale manifest must not survive a half-written regeneration.
+    let _ = std::fs::remove_file(&manifest_path);
+    let n = want.rows * want.cols;
+    for (layout, file) in &want.files {
+        // One layout at a time keeps the peak memory to about twice one
+        // file's decoded size.
+        let mut b = FileBuilder::new();
+        for k in 0..want.datasets {
+            let ds = b.create_dataset(&format!("d{k:02}"));
+            ds.with_f32_data(&dataset_values(k, n))
+                .with_shape(&[want.rows, want.cols]);
+            if layout == "deflate" {
+                ds.with_chunks(&[CHUNK.min(want.rows), CHUNK])
+                    .with_deflate(DEFLATE_LEVEL);
+            }
+        }
+        b.write(dir.join(file)).map_err(std::io::Error::other)?;
+    }
+    std::fs::write(
+        &manifest_path,
+        serde_json::to_string_pretty(want).map_err(std::io::Error::other)?,
+    )?;
+    Ok(true)
+}
+
+fn slab_offsets(seed: u64, count: usize, rows: u64, cols: u64, slab: u64) -> Vec<(u64, u64)> {
+    let mut s = seed;
+    (0..count)
+        .map(|_| {
+            let r = splitmix64(&mut s) % (rows - slab + 1);
+            let c = splitmix64(&mut s) % (cols - slab + 1);
+            (r, c)
+        })
+        .collect()
+}
+
+/// Warm the page cache by reading every byte of `path`.
+fn warm(path: &Path) -> std::io::Result<()> {
+    let mut f = std::fs::File::open(path)?;
+    std::io::copy(&mut f, &mut std::io::sink())?;
+    Ok(())
+}
+
+/// Ask the kernel to drop `path`'s pages from the page cache.
+fn evict(path: &Path) -> std::io::Result<()> {
+    use std::os::fd::AsRawFd;
+    let f = std::fs::File::open(path)?;
+    // SAFETY: plain syscall on a valid, open file descriptor.
+    let rc = unsafe { libc::posix_fadvise(f.as_raw_fd(), 0, 0, libc::POSIX_FADV_DONTNEED) };
+    if rc != 0 {
+        return Err(std::io::Error::from_raw_os_error(rc));
+    }
+    Ok(())
+}
+
+#[derive(Serialize)]
+struct Row {
+    layout: String,
+    mode: String,
+    threads: usize,
+    /// Decoded (selected) bytes read per repetition.
+    bytes: u64,
+    times_s: Vec<f64>,
+    median_s: f64,
+    mb_s: f64,
+    /// `mb_s / (threads * mb_s at threads = 1)`; null without a 1-thread row.
+    efficiency: Option<f64>,
+}
+
+struct Args {
+    dir: PathBuf,
+    datasets: u64,
+    mib: u64,
+    threads: Vec<usize>,
+    reps: usize,
+    slab: u64,
+    slabs: usize,
+    seed: u64,
+    cold: bool,
+    decode_threads: usize,
+    modes: Vec<String>,
+    layouts: Vec<String>,
+    json: Option<PathBuf>,
+}
+
+const USAGE: &str = "\
+usage: concurrent_read [--dir DIR] [--datasets N] [--mib N] [--threads 1,2,4,8,16]
+                       [--reps N] [--slab N] [--slabs N] [--seed N] [--cold]
+                       [--decode-threads N] [--modes distinct,same]
+                       [--layouts deflate,contiguous] [--json FILE]";
+
+fn parse_list<T: std::str::FromStr>(s: &str) -> Result<Vec<T>, String> {
+    s.split(',')
+        .map(|x| x.trim().parse().map_err(|_| format!("bad list item {x:?}")))
+        .collect()
+}
+
+fn parse_args() -> Result<Args, String> {
+    let mut a = Args {
+        dir: PathBuf::from("concurrent-read-data"),
+        datasets: 64,
+        mib: 64,
+        threads: vec![1, 2, 4, 8, 16],
+        reps: 3,
+        slab: 256,
+        slabs: 1024,
+        seed: 42,
+        cold: false,
+        decode_threads: 0,
+        modes: vec!["distinct".into(), "same".into()],
+        layouts: LAYOUTS.iter().map(|s| s.to_string()).collect(),
+        json: None,
+    };
+    let mut it = std::env::args().skip(1);
+    while let Some(flag) = it.next() {
+        if flag == "--cold" {
+            a.cold = true;
+            continue;
+        }
+        if flag == "-h" || flag == "--help" {
+            return Err(USAGE.into());
+        }
+        let v = it.next().ok_or(format!("{flag} needs a value\n{USAGE}"))?;
+        let num = |v: &str| {
+            v.parse::<u64>()
+                .map_err(|_| format!("{flag}: bad number {v:?}"))
+        };
+        match flag.as_str() {
+            "--dir" => a.dir = v.into(),
+            "--datasets" => a.datasets = num(&v)?,
+            "--mib" => a.mib = num(&v)?,
+            "--threads" => a.threads = parse_list(&v)?,
+            "--reps" => a.reps = num(&v)? as usize,
+            "--slab" => a.slab = num(&v)?,
+            "--slabs" => a.slabs = num(&v)? as usize,
+            "--seed" => a.seed = num(&v)?,
+            "--decode-threads" => a.decode_threads = num(&v)? as usize,
+            "--modes" => a.modes = parse_list(&v)?,
+            "--layouts" => a.layouts = parse_list(&v)?,
+            "--json" => a.json = Some(v.into()),
+            _ => return Err(format!("unknown flag {flag}\n{USAGE}")),
+        }
+    }
+    if a.datasets == 0 || a.datasets > 100 {
+        return Err("--datasets must be 1..=100".into());
+    }
+    if a.mib == 0 || a.reps == 0 || a.slabs == 0 || a.threads.contains(&0) {
+        return Err("--mib, --reps, --slabs and every --threads value must be > 0".into());
+    }
+    if a.slab == 0 || a.slab > COLS || a.slab > a.mib * ROWS_PER_MIB {
+        return Err(format!(
+            "--slab must be 1..={}",
+            COLS.min(a.mib * ROWS_PER_MIB)
+        ));
+    }
+    for m in &a.modes {
+        if m != "distinct" && m != "same" {
+            return Err(format!("unknown mode {m:?}"));
+        }
+    }
+    for l in &a.layouts {
+        if !LAYOUTS.contains(&l.as_str()) {
+            return Err(format!("unknown layout {l:?}"));
+        }
+    }
+    Ok(a)
+}
+
+/// One timed repetition: `T` threads on one shared `File`. Returns seconds.
+fn run_once(
+    path: &Path,
+    mode: &str,
+    threads: usize,
+    m: &Manifest,
+    slabs: &[(u64, u64)],
+    slab: u64,
+    verify: bool,
+) -> f64 {
+    let file = File::open(path).expect("open");
+    let barrier = Barrier::new(threads + 1); // + the spawning thread
+    let n = m.rows * m.cols;
+    // Each thread times itself from the barrier; the repetition spans the
+    // earliest start to the latest finish (timing on the spawning thread
+    // instead undercounts whenever it is scheduled after the workers ran).
+    let spans: Vec<(Instant, Instant)> = std::thread::scope(|s| {
+        let handles: Vec<_> = (0..threads)
+            .map(|t| {
+                let (file, barrier) = (&file, &barrier);
+                s.spawn(move || {
+                    barrier.wait();
+                    let start = Instant::now();
+                    match mode {
+                        "distinct" => {
+                            for k in (t as u64..m.datasets).step_by(threads) {
+                                let got = file.dataset(&format!("d{k:02}")).unwrap().read_f32();
+                                let got = got.unwrap();
+                                assert_eq!(got.len() as u64, n);
+                                if verify {
+                                    for i in [0, n / 3, n - 1] {
+                                        assert_eq!(got[i as usize], value(k, i), "d{k:02}[{i}]");
+                                    }
+                                }
+                                std::hint::black_box(got);
+                            }
+                        }
+                        _ => {
+                            let ds = file.dataset("d00").unwrap();
+                            for &(r, c) in slabs.iter().skip(t).step_by(threads) {
+                                let sel = Selection::Hyperslab {
+                                    start: vec![r, c],
+                                    stride: vec![1, 1],
+                                    count: vec![slab, slab],
+                                    block: vec![1, 1],
+                                };
+                                let got = ds.read_f32_selection(&sel).unwrap();
+                                assert_eq!(got.len() as u64, slab * slab);
+                                if verify {
+                                    let last = (r + slab - 1) * m.cols + c + slab - 1;
+                                    assert_eq!(got[0], value(0, r * m.cols + c));
+                                    assert_eq!(*got.last().unwrap(), value(0, last));
+                                }
+                                std::hint::black_box(got);
+                            }
+                        }
+                    }
+                    (start, Instant::now())
+                })
+            })
+            .collect();
+        barrier.wait();
+        handles.into_iter().map(|h| h.join().unwrap()).collect()
+    });
+    let start = spans.iter().map(|s| s.0).min().unwrap();
+    let end = spans.iter().map(|s| s.1).max().unwrap();
+    (end - start).as_secs_f64()
+}
+
+fn median(v: &[f64]) -> f64 {
+    let mut s = v.to_vec();
+    s.sort_by(f64::total_cmp);
+    s[s.len() / 2]
+}
+
+fn hostname() -> String {
+    std::fs::read_to_string("/proc/sys/kernel/hostname")
+        .map(|s| s.trim().to_string())
+        .unwrap_or_else(|_| "unknown".into())
+}
+
+fn main() {
+    let args = match parse_args() {
+        Ok(a) => a,
+        Err(e) => {
+            eprintln!("{e}");
+            std::process::exit(2);
+        }
+    };
+    if cfg!(debug_assertions) {
+        eprintln!("warning: debug build — numbers are meaningless. Use --release.");
+    }
+    if args.decode_threads > 0 {
+        rayon::ThreadPoolBuilder::new()
+            .num_threads(args.decode_threads)
+            .build_global()
+            .expect("configure rayon pool");
+    }
+
+    let manifest = manifest_for(args.datasets, args.mib);
+    let t = Instant::now();
+    match ensure_files(&args.dir, &manifest) {
+        Ok(true) => eprintln!(
+            "generated {} in {:.1} s",
+            args.dir.display(),
+            t.elapsed().as_secs_f64()
+        ),
+        Ok(false) => eprintln!("reusing {}", args.dir.display()),
+        Err(e) => {
+            eprintln!("cannot write test files in {}: {e}", args.dir.display());
+            std::process::exit(1);
+        }
+    }
+    let path_of = |layout: &str| args.dir.join(format!("{layout}.h5"));
+    let slabs = slab_offsets(
+        args.seed,
+        args.slabs,
+        manifest.rows,
+        manifest.cols,
+        args.slab,
+    );
+    let dataset_bytes = manifest.rows * manifest.cols * 4;
+
+    let mut rows: Vec<Row> = Vec::new();
+    println!("| layout | mode | threads | MB/s | efficiency | median s |");
+    println!("|---|---|---:|---:|---:|---:|");
+    for layout in &args.layouts {
+        let path = path_of(layout);
+        // Untimed pass: page cache warm (unless --cold), results checked.
+        if !args.cold {
+            warm(&path).expect("warm page cache");
+        }
+        for mode in &args.modes {
+            run_once(&path, mode, 1, &manifest, &slabs, args.slab, true);
+            let bytes = match mode.as_str() {
+                "distinct" => dataset_bytes * manifest.datasets,
+                _ => args.slab * args.slab * 4 * args.slabs as u64,
+            };
+            let mut base: Option<f64> = None;
+            for &threads in &args.threads {
+                let times: Vec<f64> = (0..args.reps)
+                    .map(|_| {
+                        if args.cold {
+                            evict(&path).expect("posix_fadvise");
+                        }
+                        run_once(&path, mode, threads, &manifest, &slabs, args.slab, false)
+                    })
+                    .collect();
+                let med = median(&times);
+                let mb_s = bytes as f64 / (1 << 20) as f64 / med;
+                if threads == 1 {
+                    base = Some(mb_s);
+                }
+                let efficiency = base.map(|b| mb_s / (threads as f64 * b));
+                println!(
+                    "| {layout} | {mode} | {threads} | {mb_s:.0} | {} | {med:.4} |",
+                    efficiency.map_or("-".into(), |e| format!("{e:.2}"))
+                );
+                rows.push(Row {
+                    layout: layout.clone(),
+                    mode: mode.clone(),
+                    threads,
+                    bytes,
+                    times_s: times,
+                    median_s: med,
+                    mb_s,
+                    efficiency,
+                });
+            }
+        }
+    }
+
+    if let Some(out) = &args.json {
+        let doc = serde_json::json!({
+            "tool": "clawhdf5",
+            "version": env!("CARGO_PKG_VERSION"),
+            "host": hostname(),
+            "cpus": std::thread::available_parallelism().map_or(0, |n| n.get()),
+            "unix_time": std::time::SystemTime::now()
+                .duration_since(std::time::UNIX_EPOCH)
+                .map_or(0, |d| d.as_secs()),
+            "cache": if args.cold { "cold (posix_fadvise DONTNEED before each repetition)" } else { "warm" },
+            "decode_threads": rayon::current_num_threads(),
+            "params": {
+                "datasets": manifest.datasets,
+                "mib": args.mib,
+                "rows": manifest.rows,
+                "cols": manifest.cols,
+                "chunk": manifest.chunk,
+                "deflate_level": manifest.deflate_level,
+                "slab": args.slab,
+                "slabs": args.slabs,
+                "seed": args.seed,
+                "reps": args.reps,
+                "dir": args.dir,
+            },
+            "results": rows,
+        });
+        std::fs::write(out, serde_json::to_string_pretty(&doc).unwrap()).expect("write json");
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn values_are_exact_in_f32() {
+        for k in [0, 7, 63] {
+            for i in [0u64, 1, 4095, 1 << 20, (1 << 24) - 1] {
+                let v = value(k, i);
+                assert_eq!(v, (v as f64) as f32);
+                assert!(v < 32768.0);
+                assert_eq!((v * 256.0).fract(), 0.0);
+            }
+        }
+    }
+
+    /// The h5py script hard-codes this vector to check its splitmix64 port.
+    #[test]
+    fn splitmix64_reference() {
+        let mut s = 42;
+        assert_eq!(splitmix64(&mut s), 0xBDD7_3226_2FEB_6E95);
+    }
+}

crates/clawhdf5-bench/src/bin/consolidation_efficiency.rs

@@ -396,7 +396,7 @@ fn run_memory_reduction_benchmark() {
     println!();
     println!(
         "{:>8}  {:>10}  {:>10}  {:>10}  {:>12}",
-        "Initial", "Remaining", "Eviction%", "Signal OK?", "BM25 Speedup"
+        "Initial", "Remaining", "Eviction%", "Signal OK?", "Records ÷"
     );
     println!("{}", "-".repeat(58));
 
@@ -440,7 +440,8 @@ fn run_memory_reduction_benchmark() {
         // Check all signal records survived
         let signal_survived = signal_ids.iter().all(|&id| engine.get_by_id(id).is_some());
 
-        // Rough speedup: BM25 scales roughly linearly with record count
+        // How many times fewer records there are. Not a measured speedup —
+        // Part 1 measures search latency before and after.
         let speedup = before_count as f64 / after_count.max(1) as f64;
 
         println!(
@@ -480,7 +481,7 @@ fn main() {
     println!("  3. Reducing search latency proportional to record reduction");
     println!();
     println!(
-        "Cycle time scales sub-linearly: 100 records ~microseconds, 100K records ~tens of ms."
+        "Cycle time grows a little faster than linearly: 100 records ~microseconds, 100K records ~tens of ms."
     );
     println!("Signal records with Correction source + high access_count survive eviction.");
 }

crates/clawhdf5-bench/src/bin/footprint_bench.rs

@@ -11,12 +11,14 @@
 //!
 //! Configuration matrix:
 //!   - Text lengths: short (50 chars), medium (200 chars), long (1000 chars)
-//!   - Embedding: 384-dim f32 (1536 bytes raw per record)
+//!   - Embedding: 384-dim, stored as float16 (the default for new stores) or
+//!     f32 with `--f32`; "raw" bytes are counted as f32 input either way
 //!   - WAL: enabled and disabled
 //!
 //! # Usage
 //! ```
-//! cargo run --release --bin footprint_bench
+//! cargo run --release --bin footprint_bench           # float16 stores
+//! cargo run --release --bin footprint_bench -- --f32  # f32 stores
 //! ```
 
 use std::time::Instant;
@@ -24,6 +26,9 @@ use std::time::Instant;
 use clawhdf5_agent::{AgentMemory, HDF5Memory, MemoryConfig, MemoryEntry};
 use tempfile::TempDir;
 
+/// `--f32`: build f32 stores instead of the library's float16 default.
+static F32: std::sync::atomic::AtomicBool = std::sync::atomic::AtomicBool::new(false);
+
 const EMBEDDING_DIM: usize = 384;
 
 // Raw bytes per record: 384 f32 embeddings + median text + overhead
@@ -152,6 +157,9 @@ fn measure_footprint(
     config.compression = compression;
     config.compression_level = if compression { 6 } else { 0 };
     config.compact_threshold = 0.0;
+    if F32.load(std::sync::atomic::Ordering::Relaxed) {
+        config.float16 = false;
+    }
 
     let mut memory = HDF5Memory::create(config).expect("HDF5Memory::create failed");
 
@@ -241,11 +249,19 @@ fn fmt_n(n: usize) -> String {
 // ---------------------------------------------------------------------------
 
 fn main() {
+    if std::env::args().skip(1).any(|a| a == "--f32") {
+        F32.store(true, std::sync::atomic::Ordering::Relaxed);
+    }
+    let stored = if F32.load(std::sync::atomic::Ordering::Relaxed) {
+        "f32 (1,536 bytes per record)"
+    } else {
+        "float16 (768 bytes per record; the default for new stores)"
+    };
     println!("=================================================================");
     println!("  ClawhDF5 Memory Footprint Benchmark");
     println!("=================================================================");
     println!();
-    println!("Embedding: 384-dim f32 = 1,536 bytes raw per record");
+    println!("Embedding: 384-dim, stored as {stored}; raw input counted as f32");
     println!("Text lengths: short=50 chars, medium=200 chars, long=1000 chars");
     println!();
 

crates/clawhdf5-bench/src/bin/longmemeval_bench.rs

@@ -57,21 +57,35 @@ mod embedder;
 
 use clawhdf5_agent::bm25::TokenFilter;
 use clawhdf5_agent::hybrid::Fusion;
-use clawhdf5_agent::{AgentMemory, HDF5Memory, MemoryConfig, MemoryEntry};
+use clawhdf5_agent::reranker::{ReRankConfig, RerankInput, rerank};
+use clawhdf5_agent::{AgentMemory, HDF5Memory, MemoryConfig, MemoryEntry, SearchResult};
 use serde::Deserialize;
 use tempfile::TempDir;
 
 const EMBEDDING_DIM: usize = 384;
 
+/// `--float16`: build every per-question store with `MemoryConfig::float16`,
+/// so embeddings are rounded to half precision as they are saved — exactly
+/// what such a store searches over.
+static FLOAT16: std::sync::atomic::AtomicBool = std::sync::atomic::AtomicBool::new(false);
+
 /// A mode's fusion, as one short string for the reports.
 fn describe(mode: Mode) -> String {
     let fusion = match mode.fusion {
         Fusion::Weighted { vector, keyword } => format!("vector_{vector:.1}_keyword_{keyword:.1}"),
         Fusion::Rrf { k } => format!("rrf_k{k:.0}"),
     };
-    match mode.tokens {
+    let tokens = match mode.tokens {
         TokenFilter::Plain => fusion,
         TokenFilter::Stemmed => format!("{fusion}_stemmed"),
+    };
+    match mode.rerank {
+        None => tokens,
+        Some(cfg) if cfg.relevance_weight == 0.0 => format!("{tokens}_rerank_metadata"),
+        Some(cfg) => format!(
+            "{tokens}_rerank_blended_hl{:.0}d",
+            cfg.temporal_half_life_secs / 86_400.0
+        ),
     }
 }
 
@@ -83,6 +97,9 @@ struct Mode {
     fusion: Fusion,
     /// How keyword tokens are normalised before indexing and querying.
     tokens: TokenFilter,
+    /// Re-rank the retrieved candidates with recency and friends, relative to
+    /// the question's own date.
+    rerank: Option<ReRankConfig>,
 }
 
 impl Mode {
@@ -91,9 +108,17 @@ impl Mode {
             label,
             fusion: Fusion::Weighted { vector, keyword },
             tokens: TokenFilter::Plain,
+            rerank: None,
         }
     }
 
+    #[cfg_attr(not(feature = "embeddings"), allow(dead_code))]
+    fn reranked(mut self, label: &'static str, rerank: ReRankConfig) -> Self {
+        self.label = label;
+        self.rerank = Some(rerank);
+        self
+    }
+
     const fn stemmed(mut self, label: &'static str) -> Self {
         self.label = label;
         self.tokens = TokenFilter::Stemmed;
@@ -121,11 +146,61 @@ const RRF: Mode = Mode {
     label: "Hybrid (reciprocal rank fusion, k=60)",
     fusion: Fusion::Rrf { k: 60.0 },
     tokens: TokenFilter::Plain,
+    rerank: None,
 };
 
 /// The same two configurations with stemmed keyword tokens, so the tokenizer's
 /// effect is isolated from everything else.
 const BM25_STEMMED: Mode = BM25_ONLY.stemmed("BM25 only, stemmed tokens");
+
+/// Re-ranking as it behaved before `relevance` was an input: the combined
+/// score was recency + authority + activation only, so the retriever's own
+/// ordering was discarded.
+#[cfg(feature = "embeddings")]
+fn hybrid_rerank_metadata_only() -> Mode {
+    HYBRID.reranked(
+        "Hybrid + rerank (metadata only, pre-fix)",
+        ReRankConfig {
+            relevance_weight: 0.0,
+            ..ReRankConfig::default()
+        },
+    )
+}
+
+/// Re-ranking as it behaves now: relevance leads, recency nudges.
+#[cfg(feature = "embeddings")]
+fn hybrid_rerank_blended() -> Mode {
+    HYBRID.reranked(
+        "Hybrid + rerank (relevance + recency)",
+        ReRankConfig::default(),
+    )
+}
+
+/// The same blend at several half-lives. Decay is `2^(-age / half_life)`, so a
+/// half-life far shorter than the gaps between memories sends every score to
+/// zero and the signal vanishes; far longer and everything scores ~1 and it
+/// vanishes the other way. The right value tracks how far apart the memories
+/// actually are.
+#[cfg(feature = "embeddings")]
+fn hybrid_rerank_half_lives() -> Vec<Mode> {
+    [
+        ("1 day", 86_400.0),
+        ("7 days", 7.0 * 86_400.0),
+        ("30 days", 30.0 * 86_400.0),
+        ("90 days", 90.0 * 86_400.0),
+    ]
+    .into_iter()
+    .map(|(label, half_life)| {
+        HYBRID.reranked(
+            Box::leak(format!("Hybrid + rerank, half-life {label}").into_boxed_str()),
+            ReRankConfig {
+                temporal_half_life_secs: half_life,
+                ..ReRankConfig::default()
+            },
+        )
+    })
+    .collect()
+}
 #[cfg(feature = "embeddings")]
 const HYBRID_STEMMED: Mode = HYBRID.stemmed("Hybrid 0.4/0.6, stemmed tokens");
 
@@ -217,6 +292,37 @@ struct Question {
     haystack_session_ids: Vec<String>,
     haystack_sessions: Vec<Vec<Turn>>,
     answer_session_ids: Vec<String>,
+    /// One timestamp per haystack session, e.g. "2023/05/25 (Thu) 20:21".
+    #[serde(default)]
+    haystack_dates: Vec<String>,
+}
+
+/// Seconds since the epoch for a LongMemEval session date, which looks like
+/// `2023/05/25 (Thu) 20:21`. Sessions are stored in chronological order, so a
+/// date that cannot be parsed falls back to its position — order is preserved
+/// even if the interval is not.
+fn session_time(date: &str, position: usize) -> f64 {
+    let stamp = |y: i64, mo: i64, d: i64, h: i64, mi: i64| -> f64 {
+        // Days since 1970-01-01 via the civil-from-days algorithm.
+        let (y, mo) = if mo <= 2 { (y - 1, mo + 12) } else { (y, mo) };
+        let era = y.div_euclid(400);
+        let yoe = y - era * 400;
+        let doy = (153 * (mo - 3) + 2) / 5 + d - 1;
+        let doe = yoe * 365 + yoe / 4 - yoe / 100 + doy;
+        let days = era * 146_097 + doe - 719_468;
+        (days * 86_400 + h * 3_600 + mi * 60) as f64
+    };
+    let parse = || -> Option<f64> {
+        let (ymd, rest) = date.split_once(' ')?;
+        let mut ymd = ymd.split('/');
+        let y = ymd.next()?.parse().ok()?;
+        let mo = ymd.next()?.parse().ok()?;
+        let d = ymd.next()?.parse().ok()?;
+        let hm = rest.rsplit(' ').next()?;
+        let (h, mi) = hm.split_once(':')?;
+        Some(stamp(y, mo, d, h.parse().ok()?, mi.parse().ok()?))
+    };
+    parse().unwrap_or(1_000_000.0 + position as f64 * 86_400.0)
 }
 
 // ---------------------------------------------------------------------------
@@ -235,11 +341,21 @@ struct Metrics {
     rr_turn: f64,
     abstention_correct: u32,
     abstention_total: u32,
+    /// Questions where the newest gold session outranked the older ones, out
+    /// of those with more than one gold session and at least one retrieved.
+    newest_gold_first: u32,
+    newest_gold_total: u32,
     latency_ns: Vec<u64>,
     count: u32,
 }
 
 impl Metrics {
+    /// `None` when no question in this bucket had multiple gold sessions.
+    fn newest_gold_first_pct(&self) -> Option<f64> {
+        (self.newest_gold_total > 0)
+            .then(|| self.newest_gold_first as f64 / self.newest_gold_total as f64 * 100.0)
+    }
+
     fn hit1_session_pct(&self) -> f64 {
         self.hit1_session as f64 / self.count.max(1) as f64 * 100.0
     }
@@ -297,6 +413,16 @@ struct EvalResult {
     hit5_turn: bool,
     hit10_turn: bool,
     rr_turn: Option<f64>,
+    /// For a question whose evidence spans several dated sessions (a
+    /// `knowledge-update`, where an earlier fact is superseded by a later
+    /// one): did the *newest* gold session outrank every older gold session
+    /// that was returned? `None` when the question has one gold session, or
+    /// when none were retrieved, so there is nothing to discriminate.
+    ///
+    /// Plain recall cannot see this. LongMemEval labels *both* the stale and
+    /// the updated session as gold, so returning either counts as a hit — yet
+    /// only one of them answers the question correctly.
+    newest_gold_first: Option<bool>,
     latency: Duration,
 }
 
@@ -310,6 +436,7 @@ fn evaluate_question(
     let mut config = MemoryConfig::new(dir.path().join("lme.h5"), "lme-bench", EMBEDDING_DIM);
     config.wal_enabled = false;
     config.compact_threshold = 0.0;
+    config.float16 = FLOAT16.load(std::sync::atomic::Ordering::Relaxed);
 
     let mut memory = HDF5Memory::create(config).expect("failed to create HDF5Memory");
     memory.set_token_filter(mode.tokens);
@@ -317,15 +444,21 @@ fn evaluate_question(
     // Build MemoryEntry list from all haystack sessions
     let mut entries: Vec<MemoryEntry> = Vec::new();
     let mut turn_has_answer: Vec<bool> = Vec::new();
-    let mut ts = 1_000_000.0f64;
-
     for (sess_idx, session) in q.haystack_sessions.iter().enumerate() {
         let sess_id = q
             .haystack_session_ids
             .get(sess_idx)
             .map(String::as_str)
             .unwrap_or("unknown");
-        for turn in session {
+        // Real session dates, not a synthetic counter: anything that decays
+        // with age needs true intervals, not just the right order.
+        let session_start = q
+            .haystack_dates
+            .get(sess_idx)
+            .map_or(sess_idx as f64 * 86_400.0, |d| session_time(d, sess_idx));
+        for (turn_idx, turn) in session.iter().enumerate() {
+            // Spread a session's turns over the minutes following its start.
+            let ts = session_start + turn_idx as f64 * 60.0;
             entries.push(MemoryEntry {
                 chunk: turn.content.clone(),
                 embedding: embedding_for(embeddings, &turn.content),
@@ -339,7 +472,6 @@ fn evaluate_question(
                 },
             });
             turn_has_answer.push(turn.has_answer);
-            ts += 1.0;
         }
     }
 
@@ -356,11 +488,87 @@ fn evaluate_question(
     // Set of session IDs that contain the answer
     let answer_sess_set: HashSet<&str> = q.answer_session_ids.iter().map(String::as_str).collect();
 
+    // When each gold session was recorded, so "newest" is by date rather than
+    // by position (the two agree in this dataset, but the metric should not
+    // depend on that).
+    let gold_times: HashMap<&str, f64> = q
+        .haystack_session_ids
+        .iter()
+        .enumerate()
+        .filter(|(_, sid)| answer_sess_set.contains(sid.as_str()))
+        .map(|(i, sid)| {
+            let t = q
+                .haystack_dates
+                .get(i)
+                .map_or(i as f64 * 86_400.0, |d| session_time(d, i));
+            (sid.as_str(), t)
+        })
+        .collect();
+
     let query_emb = embedding_for(embeddings, &q.question);
     let t0 = Instant::now();
-    let results = memory.hybrid_search_with(&query_emb, &q.question, mode.fusion, top_k);
+    // Re-ranking only reorders; it needs a candidate pool larger than `top_k`
+    // to have anything to promote.
+    let pool = if mode.rerank.is_some() {
+        top_k * 4
+    } else {
+        top_k
+    };
+    let mut results = memory.hybrid_search_with(&query_emb, &q.question, mode.fusion, pool);
+    if let Some(config) = mode.rerank {
+        // "Now" is the moment the question was asked, so decay measures how
+        // stale each memory was at that point.
+        let now = session_time(&q.question_date, q.haystack_sessions.len());
+        let inputs: Vec<RerankInput> = results
+            .iter()
+            .map(|r| RerankInput {
+                index: r.index,
+                timestamp: r.timestamp,
+                source_channel: r.source_channel.clone(),
+                raw_activation: r.activation,
+                relevance: r.score,
+            })
+            .collect();
+        let order: Vec<usize> = rerank(&inputs, &config, now)
+            .into_iter()
+            .map(|r| r.index)
+            .collect();
+        let by_index: HashMap<usize, SearchResult> =
+            results.into_iter().map(|r| (r.index, r)).collect();
+        results = order
+            .into_iter()
+            .filter_map(|i| by_index.get(&i).cloned())
+            .collect();
+    }
+    results.truncate(top_k);
     let latency = t0.elapsed();
 
+    // Rank of the best-placed result from each gold session.
+    let mut first_rank: HashMap<&str, usize> = HashMap::new();
+    for (rank, result) in results.iter().enumerate() {
+        let sid = memory.cache.session_ids[result.index].as_str();
+        if let Some((gold_sid, _)) = gold_times.get_key_value(sid) {
+            first_rank.entry(gold_sid).or_insert(rank);
+        }
+    }
+    let newest_gold_first = if gold_times.len() < 2 || first_rank.is_empty() {
+        None
+    } else {
+        // The newest gold session must be retrieved, and no older gold session
+        // may outrank it.
+        let newest = gold_times
+            .iter()
+            .max_by(|a, b| a.1.total_cmp(b.1))
+            .map(|(sid, _)| *sid)
+            .expect("at least two gold sessions");
+        Some(match first_rank.get(newest) {
+            Some(&newest_rank) => first_rank
+                .iter()
+                .all(|(sid, &rank)| *sid == newest || rank > newest_rank),
+            None => false,
+        })
+    };
+
     // Session-level recall
     let mut hit1_session = false;
     let mut hit5_session = false;
@@ -415,6 +623,7 @@ fn evaluate_question(
         hit5_turn,
         hit10_turn,
         rr_turn,
+        newest_gold_first,
         latency,
     }
 }
@@ -566,6 +775,24 @@ fn print_report(
     );
     println!();
 
+    if let Some(pct) = overall.newest_gold_first_pct() {
+        println!(
+            "## Recency Discrimination  (n={})",
+            overall.newest_gold_total
+        );
+        println!(
+            "  Newest gold session ranked first: {}/{} ({pct:.1}%)",
+            overall.newest_gold_first, overall.newest_gold_total
+        );
+        println!(
+            "  Questions whose evidence spans several dated sessions — a fact and\n  \
+             its later correction. Both sessions are labelled gold, so recall\n  \
+             scores either as a hit; this asks whether the *current* one came\n  \
+             first. A retriever with no sense of time scores near chance."
+        );
+        println!();
+    }
+
     if overall.abstention_total > 0 {
         println!("## Abstention Accuracy");
         println!(
@@ -679,6 +906,14 @@ fn print_report(
     } else {
         println!("  \"abstention_accuracy\": null,");
     }
+    match overall.newest_gold_first_pct() {
+        Some(pct) => println!(
+            "  \"newest_gold_first\": {:.4}, \"newest_gold_n\": {},",
+            pct / 100.0,
+            overall.newest_gold_total
+        ),
+        None => println!("  \"newest_gold_first\": null,"),
+    }
     println!("  \"latency_us\": {{");
     println!(
         "    \"avg\": {:.1}, \"p50\": {:.1}, \"p95\": {:.1}, \"p99\": {:.1}",
@@ -701,6 +936,8 @@ fn main() {
     let mut limit: Option<usize> = None;
     let mut weights_dir: Option<String> = None;
     let mut sweep = false;
+    #[cfg_attr(not(feature = "embeddings"), allow(unused_mut, unused_variables))]
+    let mut rerank_sweep = false;
     let mut args = std::env::args().skip(1);
     while let Some(arg) = args.next() {
         match arg.as_str() {
@@ -709,6 +946,20 @@ fn main() {
                 limit = Some(v.parse().expect("--limit must be a positive integer"));
             }
             "--sweep" => sweep = true,
+            "--float16" => {
+                FLOAT16.store(true, std::sync::atomic::Ordering::Relaxed);
+                eprintln!("Stores use MemoryConfig::float16 (half-precision embeddings)");
+            }
+            "--rerank-sweep" => {
+                // Re-ranking needs the vector stage to have candidates worth
+                // reordering, so this is an embeddings-only comparison.
+                #[cfg(feature = "embeddings")]
+                {
+                    rerank_sweep = true;
+                }
+                #[cfg(not(feature = "embeddings"))]
+                eprintln!("warning: --rerank-sweep needs --features embeddings; ignoring");
+            }
             "--embeddings" => {
                 weights_dir = Some(args.next().expect("--embeddings needs a directory"));
             }
@@ -727,6 +978,12 @@ fn main() {
                               BM25-only, vector-only, and hybrid separately. Requires\n\
                               --features embeddings; without it the vector stage is\n\
                               inert and only the BM25 row is produced.\n\
+                     --rerank-sweep\n\
+                              compare re-ranking off, metadata-only (the old\n\
+                              behaviour) and blended at several half-lives.\n\
+                     --float16\n\
+                              build each store with MemoryConfig::float16, to\n\
+                              compare retrieval on half-precision embeddings.\n\
                      --sweep  instead of the three named modes, sweep vector_weight\n\
                               from 0.0 to 1.0 in 0.1 steps. The 0.7/0.3 default was\n\
                               never searched; this is what searches it."
@@ -792,6 +1049,10 @@ fn main() {
         {
             if sweep {
                 sweep_modes()
+            } else if rerank_sweep {
+                let mut modes = vec![HYBRID, hybrid_rerank_metadata_only()];
+                modes.extend(hybrid_rerank_half_lives());
+                modes
             } else {
                 vec![
                     BM25_ONLY,
@@ -800,6 +1061,8 @@ fn main() {
                     RRF,
                     BM25_STEMMED,
                     HYBRID_STEMMED,
+                    hybrid_rerank_metadata_only(),
+                    hybrid_rerank_blended(),
                 ]
             }
         }
@@ -916,6 +1179,14 @@ fn run_mode(
                 entry.rr_turn += rr;
                 overall.rr_turn += rr;
             }
+            if let Some(newest_first) = result.newest_gold_first {
+                entry.newest_gold_total += 1;
+                overall.newest_gold_total += 1;
+                if newest_first {
+                    entry.newest_gold_first += 1;
+                    overall.newest_gold_first += 1;
+                }
+            }
 
             let ns = result.latency.as_nanos() as u64;
             entry.latency_ns.push(ns);
@@ -927,3 +1198,30 @@ fn run_mode(
     eprintln!();
     print_report(&overall, &by_type, profile, mode);
 }
+
+#[cfg(test)]
+mod tests {
+    use super::session_time;
+
+    #[test]
+    fn session_dates_parse_to_the_right_instant() {
+        // Reference values from Python's datetime, UTC.
+        for (date, expected) in [
+            ("2023/05/25 (Thu) 20:21", 1_685_046_060.0),
+            ("1970/01/01 (Thu) 00:00", 0.0),
+            ("2000/02/29 (Tue) 12:00", 951_825_600.0),
+            ("2023/12/31 (Sun) 23:59", 1_704_067_140.0),
+            ("2024/03/01 (Fri) 00:00", 1_709_251_200.0),
+        ] {
+            assert_eq!(session_time(date, 0), expected, "{date}");
+        }
+    }
+
+    #[test]
+    fn unparseable_dates_fall_back_to_position_order() {
+        let a = session_time("not a date", 0);
+        let b = session_time("", 1);
+        let c = session_time("2023/13/99 (???) 99:99", 2);
+        assert!(a < b && b < c, "fallback must preserve session order");
+    }
+}

crates/clawhdf5-bench/src/bin/read_harness.rs

@@ -7,15 +7,19 @@
 //! ```text
 //! cargo run --release -p clawhdf5-bench --bin read_harness
 //! cargo run --release -p clawhdf5-bench --bin read_harness -- --large   # 512 MB
+//! cargo run --release -p clawhdf5-bench --bin read_harness -- --v18      # HDF5 1.8 format
+//! cargo run --release -p clawhdf5-bench --bin read_harness -- --chunk 32 # 32 x 32 chunks
 //! ```
+//!
+//! `--v18` writes the file with `libver_bounds(V18, V18)` (version-1 B-tree
+//! chunk indexes) instead of the default 1.10 format (Fixed Array indexes
+//! here), to compare the two.
 
 use std::time::{Duration, Instant};
 
-use clawhdf5::{File, FileBuilder};
+use clawhdf5::{File, FileBuilder, LibVer};
 use clawhdf5_format::selection::Selection;
 
-const CHUNK: u64 = 256;
-
 struct Layout {
     name: &'static str,
     chunked: bool,
@@ -46,16 +50,19 @@ fn value(row: u64, col: u64) -> f64 {
     (row * 100_003 + col) as f64 * 0.5
 }
 
-fn write_file(path: &std::path::Path, rows: u64, cols: u64) {
+fn write_file(path: &std::path::Path, rows: u64, cols: u64, chunk: u64, v18: bool) {
     let data: Vec<f64> = (0..rows)
         .flat_map(|r| (0..cols).map(move |c| value(r, c)))
         .collect();
     let mut builder = FileBuilder::new();
+    if v18 {
+        builder.libver_bounds(LibVer::V18, LibVer::V18);
+    }
     for (i, layout) in LAYOUTS.iter().enumerate() {
         let ds = builder.create_dataset(&format!("d{i}"));
         ds.with_f64_data(&data).with_shape(&[rows, cols]);
         if layout.chunked {
-            ds.with_chunks(&[CHUNK, CHUNK]);
+            ds.with_chunks(&[chunk, chunk]);
         }
         if layout.deflate {
             ds.with_deflate(4);
@@ -91,7 +98,14 @@ fn slab(start: [u64; 2], count: [u64; 2]) -> Selection {
 }
 
 fn main() {
-    let large = std::env::args().any(|a| a == "--large");
+    let args: Vec<String> = std::env::args().collect();
+    let large = args.iter().any(|a| a == "--large");
+    let v18 = args.iter().any(|a| a == "--v18");
+    let chunk: u64 = args
+        .iter()
+        .position(|a| a == "--chunk")
+        .and_then(|i| args.get(i + 1))
+        .map_or(256, |c| c.parse().expect("--chunk N"));
     let (rows, cols) = if large { (8192, 8192) } else { (4096, 2048) };
     let total_mb = (rows * cols * 8) as f64 / (1 << 20) as f64;
     if cfg!(debug_assertions) {
@@ -100,12 +114,21 @@ fn main() {
 
     let dir = tempfile::TempDir::new().unwrap();
     let path = dir.path().join("read_harness.h5");
-    write_file(&path, rows, cols);
-    let file_mb = std::fs::metadata(&path).unwrap().len() as f64 / (1 << 20) as f64;
+    let t = Instant::now();
+    write_file(&path, rows, cols, chunk, v18);
+    let write_ms = t.elapsed().as_secs_f64() * 1e3;
+    let file_bytes = std::fs::metadata(&path).unwrap().len();
+    let file_mb = file_bytes as f64 / (1 << 20) as f64;
 
     println!("## Read harness");
     println!(
-        "\n{rows} x {cols} f64 ({total_mb:.0} MB per dataset), chunks {CHUNK} x {CHUNK}, file {file_mb:.0} MB\n"
+        "\n{rows} x {cols} f64 ({total_mb:.0} MB per dataset), chunks {chunk} x {chunk}, \
+         format {}, file {file_mb:.0} MB ({file_bytes} bytes), written in {write_ms:.0} ms\n",
+        if v18 {
+            "1.8 (v1 B-tree)"
+        } else {
+            "1.10 (default)"
+        }
     );
 
     // (label, selection, elements selected)

crates/clawhdf5-bench/src/bin/search_harness.rs

@@ -19,12 +19,15 @@
 //! cargo run --release -p clawhdf5-bench --bin search_harness -- --full  # + 100K
 //! cargo run --release -p clawhdf5-bench --bin search_harness -- --json out.json
 //! cargo run --release -p clawhdf5-bench --bin search_harness -- --ann-only --uniform
+//! cargo run --release -p clawhdf5-bench --bin search_harness -- --float16-study --full
+//! cargo run --release -p clawhdf5-bench --bin search_harness -- --options-study --full
+//! cargo run --release -p clawhdf5-bench --bin search_harness -- --signing-study --full
 //! ```
 
 use std::time::{Duration, Instant};
 
 use clawhdf5_agent::{AgentMemory, HDF5Memory, MemoryConfig, MemoryEntry};
-use clawhdf5_ann::{DistanceMetric, HnswIndex};
+use clawhdf5_ann::{DistanceMetric, HnswIndex, Storage};
 
 const DIM: usize = 384;
 const K: usize = 10;
@@ -84,6 +87,25 @@ struct Dataset {
 /// that appears only on clustered data points at graph connectivity.
 static UNIFORM: std::sync::atomic::AtomicBool = std::sync::atomic::AtomicBool::new(false);
 
+/// `--int8`: build the HNSW index over int8-quantised vectors (a quarter of
+/// the memory) instead of f32, to price the recall it costs.
+static INT8: std::sync::atomic::AtomicBool = std::sync::atomic::AtomicBool::new(false);
+
+/// `--f16-first`: in `--float16-study`, run the float16 store first.
+static F16_FIRST: std::sync::atomic::AtomicBool = std::sync::atomic::AtomicBool::new(false);
+
+/// `--rerank`: re-score the candidate pool against the exact vectors before
+/// taking the top K.
+static RERANK: std::sync::atomic::AtomicBool = std::sync::atomic::AtomicBool::new(false);
+
+fn storage() -> Storage {
+    if INT8.load(std::sync::atomic::Ordering::Relaxed) {
+        Storage::Int8
+    } else {
+        Storage::Float32
+    }
+}
+
 fn make_dataset(n: usize, seed: u64) -> Dataset {
     let mut rng = Rng(seed);
     if UNIFORM.load(std::sync::atomic::Ordering::Relaxed) {
@@ -169,6 +191,11 @@ fn text_for(cluster: usize, i: usize, rng: &mut Rng) -> String {
 // Measurement helpers
 // ---------------------------------------------------------------------------
 
+/// Exact cosine distance between unit-length vectors.
+fn exact_dist(a: &[f32], b: &[f32]) -> f32 {
+    1.0 - a.iter().zip(b).map(|(x, y)| x * y).sum::<f32>()
+}
+
 fn exact_top_k(vectors: &[Vec<f32>], query: &[f32], k: usize) -> Vec<usize> {
     // Vectors are unit length, so cosine order == dot-product order.
     let mut scored: Vec<(usize, f32)> = vectors
@@ -198,6 +225,83 @@ fn summarize(mut samples: Vec<Duration>) -> Latency {
     }
 }
 
+/// Counts live heap bytes, so a structure's cost can be measured by
+/// difference.
+///
+/// RSS cannot do this from inside one process: freeing a large structure
+/// returns its pages to the allocator's pool rather than to the OS, so
+/// allocating the next one shows no change. Measured that way, a store that
+/// holds the corpus twice and one that holds it once look identical.
+struct CountingAllocator;
+
+static LIVE_BYTES: std::sync::atomic::AtomicI64 = std::sync::atomic::AtomicI64::new(0);
+
+/// High-water mark of [`LIVE_BYTES`] since it was last reset.
+///
+/// Live bytes at a checkpoint cannot see a buffer that was allocated and
+/// freed in between, and that is exactly the shape of a transient copy —
+/// which still has to fit in memory while it exists.
+static PEAK_BYTES: std::sync::atomic::AtomicI64 = std::sync::atomic::AtomicI64::new(0);
+
+fn note_peak(live: i64) {
+    PEAK_BYTES.fetch_max(live, std::sync::atomic::Ordering::Relaxed);
+}
+
+// SAFETY: every method forwards to the system allocator with the same layout
+// it was given, and only adds bookkeeping around it.
+unsafe impl std::alloc::GlobalAlloc for CountingAllocator {
+    unsafe fn alloc(&self, layout: std::alloc::Layout) -> *mut u8 {
+        let ptr = unsafe { std::alloc::System.alloc(layout) };
+        if !ptr.is_null() {
+            let live = LIVE_BYTES
+                .fetch_add(layout.size() as i64, std::sync::atomic::Ordering::Relaxed)
+                + layout.size() as i64;
+            note_peak(live);
+        }
+        ptr
+    }
+
+    unsafe fn dealloc(&self, ptr: *mut u8, layout: std::alloc::Layout) {
+        LIVE_BYTES.fetch_sub(layout.size() as i64, std::sync::atomic::Ordering::Relaxed);
+        unsafe { std::alloc::System.dealloc(ptr, layout) }
+    }
+
+    unsafe fn realloc(&self, ptr: *mut u8, layout: std::alloc::Layout, new_size: usize) -> *mut u8 {
+        let new_ptr = unsafe { std::alloc::System.realloc(ptr, layout, new_size) };
+        if !new_ptr.is_null() {
+            let delta = new_size as i64 - layout.size() as i64;
+            let live = LIVE_BYTES.fetch_add(delta, std::sync::atomic::Ordering::Relaxed) + delta;
+            note_peak(live);
+        }
+        new_ptr
+    }
+}
+
+#[global_allocator]
+static ALLOCATOR: CountingAllocator = CountingAllocator;
+
+/// Live heap bytes right now.
+fn heap_bytes() -> u64 {
+    LIVE_BYTES.load(std::sync::atomic::Ordering::Relaxed).max(0) as u64
+}
+
+/// Start watching for a new high-water mark from the current live total.
+fn reset_peak() {
+    PEAK_BYTES.store(
+        LIVE_BYTES.load(std::sync::atomic::Ordering::Relaxed),
+        std::sync::atomic::Ordering::Relaxed,
+    );
+}
+
+/// The highest live total seen since [`reset_peak`].
+fn peak_bytes() -> u64 {
+    PEAK_BYTES.load(std::sync::atomic::Ordering::Relaxed).max(0) as u64
+}
+
+fn mib(bytes: u64) -> f64 {
+    bytes as f64 / (1 << 20) as f64
+}
+
 fn micros(d: Duration) -> f64 {
     d.as_secs_f64() * 1e6
 }
@@ -219,11 +323,12 @@ fn bench_ann(n: usize, json: &mut Vec<serde_json::Value>) {
         .collect();
 
     let started = Instant::now();
-    let index = HnswIndex::build_with_metric(
+    let index = HnswIndex::build_with(
         &data.vectors,
         HNSW_M,
         HNSW_EF_CONSTRUCTION,
         DistanceMetric::Cosine,
+        storage(),
     );
     let build = started.elapsed();
 
@@ -240,7 +345,8 @@ fn bench_ann(n: usize, json: &mut Vec<serde_json::Value>) {
     );
 
     println!(
-        "\n### HNSW, N = {n}, dim = {DIM}, M = {HNSW_M}, ef_construction = {HNSW_EF_CONSTRUCTION}\n"
+        "\n### HNSW, N = {n}, dim = {DIM}, M = {HNSW_M}, ef_construction = {HNSW_EF_CONSTRUCTION}, storage = {:?}\n",
+        index.storage()
     );
     println!(
         "build: {:.1} ms ({:.0} vectors/s) · exact scan: {:.0} QPS, p50 {:.0} µs\n",
@@ -251,12 +357,26 @@ fn bench_ann(n: usize, json: &mut Vec<serde_json::Value>) {
     );
     println!("| ef | recall@{K} | QPS | p50 µs | p99 µs |");
     println!("|---:|---:|---:|---:|---:|");
+    // With a quantised index the distances it returns are approximate, so
+    // the candidates are re-scored against the exact vectors the caller
+    // already holds (in the agent, the embedding cache) before taking the
+    // top K. `--rerank` prices that: it costs one exact distance per
+    // candidate and is what decides whether int8 is usable.
+    let rerank = RERANK.load(std::sync::atomic::Ordering::Relaxed);
+    let pool = if rerank { K * 4 } else { K };
     for ef in EF_VALUES {
         let mut hits = 0usize;
         let mut samples = Vec::with_capacity(data.queries.len());
         for (q, want) in data.queries.iter().zip(&truth) {
             let t = Instant::now();
-            let got = index.search(q, K, ef);
+            let mut got = index.search(q, pool, ef.max(pool));
+            if rerank {
+                for cand in &mut got {
+                    cand.1 = exact_dist(&data.vectors[cand.0], q);
+                }
+                got.select_nth_unstable_by(K - 1, |a, b| a.1.total_cmp(&b.1));
+                got.truncate(K);
+            }
             samples.push(t.elapsed());
             hits += got.iter().filter(|(id, _)| want.contains(id)).count();
         }
@@ -369,6 +489,394 @@ fn bench_end_to_end(n: usize, json: &mut Vec<serde_json::Value>) {
     }));
 }
 
+// ---------------------------------------------------------------------------
+// Signing study: what does an Ed25519-signed checkpoint cost?
+// ---------------------------------------------------------------------------
+
+/// `--signing-study`: checkpoint time unsigned vs signed, `verify` time, and
+/// the file-size cost of the stored per-record hashes. Default store
+/// settings (float16, int8 index). Medians of five checkpoints / three
+/// verifies.
+fn signing_study(n: usize) {
+    use clawhdf5_agent::signing::SigningKey;
+    let data = make_dataset(n, 0x516 ^ n as u64);
+    let mut rng = Rng(9);
+    let entries: Vec<MemoryEntry> = data
+        .vectors
+        .iter()
+        .enumerate()
+        .map(|(i, v)| MemoryEntry {
+            chunk: text_for(data.cluster_of[i], i, &mut rng),
+            embedding: v.clone(),
+            source_channel: "bench".into(),
+            timestamp: i as f64,
+            session_id: format!("s{}", i % 50),
+            tags: format!("t{i}"),
+        })
+        .collect();
+    let dir = tempfile::tempdir().unwrap();
+    let path = dir.path().join("sign.h5");
+    let mut mem = HDF5Memory::create(MemoryConfig::new(path.clone(), "bench", DIM)).unwrap();
+    mem.save_batch(entries).unwrap();
+    std::hint::black_box(mem.hybrid_search(&data.queries[0], "", 1.0, 0.0, K));
+
+    let median = |mut v: Vec<Duration>| {
+        v.sort();
+        v[v.len() / 2]
+    };
+    let checkpoint = |mem: &mut HDF5Memory| {
+        median(
+            (0..5)
+                .map(|_| {
+                    let t = Instant::now();
+                    mem.flush_wal().unwrap();
+                    t.elapsed()
+                })
+                .collect(),
+        )
+    };
+    let unsigned = checkpoint(&mut mem);
+    let unsigned_bytes = std::fs::metadata(&path).unwrap().len();
+    let key = SigningKey::from_bytes(&[7; 32]);
+    mem.set_signing_key(key.clone());
+    let signed = checkpoint(&mut mem);
+    let signed_bytes = std::fs::metadata(&path).unwrap().len();
+    drop(mem);
+    let vk = key.verifying_key();
+    let verify = median(
+        (0..3)
+            .map(|_| {
+                let t = Instant::now();
+                let r = HDF5Memory::verify(&path, &vk).unwrap();
+                let d = t.elapsed();
+                assert!(r.is_valid());
+                d
+            })
+            .collect(),
+    );
+    println!(
+        "| {n} | {:.1} | {:.1} | {:+.1} | {:.1} | {:+.2} |",
+        millis(unsigned),
+        millis(signed),
+        millis(signed) - millis(unsigned),
+        millis(verify),
+        (signed_bytes as f64 - unsigned_bytes as f64) / (1024.0 * 1024.0),
+    );
+}
+
+// ---------------------------------------------------------------------------
+// Search options study: source filters, re-ranking, confidence rejection
+// ---------------------------------------------------------------------------
+
+/// `--options-study`: what `HDF5Memory::search`'s options cost and whether a
+/// filtered search finds the right records. Filters keep 50%, 10% or 1% of
+/// the store at random, or two whole clusters away from the query (the case
+/// the index cannot serve, which falls back to an exact scan). Recall is
+/// vector-only against an exact scan of the allowed records; latency is full
+/// hybrid search. Hebbian boosting is off.
+fn options_study(n: usize) {
+    use clawhdf5_agent::SearchOptions;
+    use clawhdf5_agent::confidence::ConfidenceConfig;
+    use clawhdf5_agent::hybrid::Fusion;
+    use clawhdf5_agent::reranker::ReRankConfig;
+
+    let data = make_dataset(n, 0x0B7 ^ n as u64);
+    let n_clusters = data.cluster_of.iter().max().map_or(1, |m| m + 1);
+    let mut rng = Rng(5);
+    let bucket_of: Vec<usize> = (0..n).map(|_| rng.below(100)).collect();
+    let bucket = &bucket_of;
+    let query_texts: Vec<String> = data
+        .query_cluster
+        .iter()
+        .enumerate()
+        .map(|(i, c)| text_for(*c, i, &mut rng))
+        .collect();
+    let exact_top = |q: &[f32], allowed: &dyn Fn(usize) -> bool| -> Vec<usize> {
+        let mut s: Vec<(usize, f32)> = (0..n)
+            .filter(|&i| allowed(i))
+            .map(|i| (i, data.vectors[i].iter().zip(q).map(|(a, b)| a * b).sum()))
+            .collect();
+        s.sort_by(|a, b| b.1.total_cmp(&a.1).then(a.0.cmp(&b.0)));
+        s.into_iter().take(K).map(|(i, _)| i).collect()
+    };
+
+    // Two stores: channel = random bucket, and channel = cluster.
+    let dir = tempfile::tempdir().unwrap();
+    let mut stores = Vec::new();
+    for by_cluster in [false, true] {
+        let mut rng = Rng(3);
+        let entries: Vec<MemoryEntry> = data
+            .vectors
+            .iter()
+            .enumerate()
+            .map(|(i, v)| MemoryEntry {
+                chunk: text_for(data.cluster_of[i], i, &mut rng),
+                embedding: v.clone(),
+                source_channel: if by_cluster {
+                    format!("c{}", data.cluster_of[i])
+                } else {
+                    format!("b{}", bucket[i])
+                },
+                timestamp: i as f64,
+                session_id: format!("s{}", i % 50),
+                tags: format!("t{i}"),
+            })
+            .collect();
+        let mut config = MemoryConfig::new(
+            dir.path().join(format!("opt_{by_cluster}.h5")),
+            "bench",
+            DIM,
+        );
+        config.hebbian_boost = 0.0;
+        let mut mem = HDF5Memory::create(config).unwrap();
+        mem.save_batch(entries).unwrap();
+        std::hint::black_box(mem.search(&data.queries[0], "", &SearchOptions::new(K)));
+        stores.push(mem);
+    }
+
+    let vector_only = SearchOptions::new(K).with_fusion(Fusion::Weighted {
+        vector: 1.0,
+        keyword: 0.0,
+    });
+    // (label, store, channels for query i, allowed(i, record))
+    type Case<'a> = (
+        String,
+        usize,
+        Box<dyn Fn(usize) -> Option<Vec<String>> + 'a>,
+        Box<dyn Fn(usize, usize) -> bool + 'a>,
+    );
+    let mut cases: Vec<Case> = vec![(
+        "no filter".into(),
+        0,
+        Box::new(|_| None),
+        Box::new(|_, _| true),
+    )];
+    for pct in [50usize, 10, 1] {
+        cases.push((
+            format!("random {pct}%"),
+            0,
+            Box::new(move |_| Some((0..pct).map(|b| format!("b{b}")).collect())),
+            Box::new(move |_, i| bucket[i] < pct),
+        ));
+    }
+    let d = &data;
+    let away = move |qi: usize| {
+        let qc = d.query_cluster[qi];
+        [
+            (qc + n_clusters / 3) % n_clusters,
+            (qc + 2 * n_clusters / 3) % n_clusters,
+        ]
+    };
+    cases.push((
+        "2 clusters away from the query".into(),
+        1,
+        Box::new(move |qi| Some(away(qi).iter().map(|c| format!("c{c}")).collect())),
+        Box::new(move |qi, i| away(qi).contains(&d.cluster_of[i])),
+    ));
+
+    for (label, store, channels, allowed) in &cases {
+        let mem = &mut stores[*store];
+        let mut hits = 0;
+        let mut kept = 0;
+        for (qi, q) in data.queries.iter().enumerate() {
+            let mut opts = vector_only.clone();
+            opts.source_channels = channels(qi);
+            let got = mem.search(q, "", &opts);
+            let want = exact_top(q, &|i| allowed(qi, i));
+            kept += want.len();
+            hits += got.iter().filter(|r| want.contains(&r.index)).count();
+        }
+        let latency = summarize(
+            (0..N_QUERIES)
+                .map(|qi| {
+                    let mut opts = SearchOptions::new(K);
+                    opts.source_channels = channels(qi);
+                    let t = Instant::now();
+                    std::hint::black_box(mem.search(&data.queries[qi], &query_texts[qi], &opts));
+                    t.elapsed()
+                })
+                .collect(),
+        );
+        println!(
+            "| {n} | {label} | {:.4} | {:.3} | {:.3} |",
+            hits as f64 / kept.max(1) as f64,
+            millis(latency.p50),
+            millis(latency.p99),
+        );
+    }
+
+    let mem = &mut stores[0];
+    for (label, opts) in [
+        (
+            "re-rank",
+            SearchOptions::new(K).with_rerank(ReRankConfig::default()),
+        ),
+        (
+            "re-rank + confidence",
+            SearchOptions::new(K)
+                .with_rerank(ReRankConfig::default())
+                .with_confidence(ConfidenceConfig::default()),
+        ),
+    ] {
+        let latency = summarize(
+            (0..N_QUERIES)
+                .map(|qi| {
+                    let t = Instant::now();
+                    std::hint::black_box(mem.search(&data.queries[qi], &query_texts[qi], &opts));
+                    t.elapsed()
+                })
+                .collect(),
+        );
+        println!(
+            "| {n} | {label} | — | {:.3} | {:.3} |",
+            millis(latency.p50),
+            millis(latency.p99)
+        );
+    }
+}
+
+// ---------------------------------------------------------------------------
+// float16 study: what does half-precision embedding storage cost?
+// ---------------------------------------------------------------------------
+
+/// `--float16-study`: the same data in an `f32` store and a `float16` store.
+/// Reports file size, checkpoint and open time, vector-search recall@10
+/// against an exact scan of the *original* f32 vectors, how often the two
+/// stores return the same top 10, and `hybrid_search` latency. Hebbian
+/// boosting is off, so every query sees the same store.
+fn float16_study(n: usize) {
+    let data = make_dataset(n, 0xF16 ^ n as u64);
+    let mut rng = Rng(11);
+    let query_texts: Vec<String> = data
+        .query_cluster
+        .iter()
+        .enumerate()
+        .map(|(i, c)| text_for(*c, i, &mut rng))
+        .collect();
+
+    // Exact top K by cosine (the vectors are unit length) on the f32 inputs.
+    let exact: Vec<Vec<usize>> = data
+        .queries
+        .iter()
+        .map(|q| {
+            let mut scored: Vec<(usize, f32)> = data
+                .vectors
+                .iter()
+                .enumerate()
+                .map(|(i, v)| (i, v.iter().zip(q).map(|(a, b)| a * b).sum()))
+                .collect();
+            scored.sort_by(|a, b| b.1.total_cmp(&a.1).then(a.0.cmp(&b.0)));
+            scored.into_iter().take(K).map(|(i, _)| i).collect()
+        })
+        .collect();
+
+    let dir = tempfile::tempdir().unwrap();
+    let mut per_variant: Vec<(bool, Vec<Vec<usize>>)> = Vec::new();
+    // `--f16-first` swaps the order, to check the numbers do not depend on
+    // which store runs first (page cache, allocator, CPU frequency).
+    let order = if F16_FIRST.load(std::sync::atomic::Ordering::Relaxed) {
+        [true, false]
+    } else {
+        [false, true]
+    };
+    for float16 in order {
+        let path = dir.path().join(format!("f16study_{float16}.h5"));
+        let mut rng = Rng(3);
+        let entries: Vec<MemoryEntry> = data
+            .vectors
+            .iter()
+            .enumerate()
+            .map(|(i, v)| MemoryEntry {
+                chunk: text_for(data.cluster_of[i], i, &mut rng),
+                embedding: v.clone(),
+                source_channel: "bench".into(),
+                timestamp: i as f64,
+                session_id: format!("s{}", i % 50),
+                tags: format!("t{i}"),
+            })
+            .collect();
+        let mut config = MemoryConfig::new(path.clone(), "bench", DIM);
+        config.float16 = float16;
+        config.hebbian_boost = 0.0;
+        let mut mem = HDF5Memory::create(config).unwrap();
+        mem.save_batch(entries).unwrap();
+        // Build the indexes, then time a checkpoint that writes everything.
+        std::hint::black_box(mem.hybrid_search(&data.queries[0], "", 1.0, 0.0, K));
+        let t = Instant::now();
+        mem.flush_wal().unwrap();
+        let checkpoint = t.elapsed();
+        drop(mem);
+        let file_bytes = std::fs::metadata(&path).unwrap().len();
+
+        // Median of three opens.
+        let mut opens: Vec<Duration> = (0..3)
+            .map(|_| {
+                let t = Instant::now();
+                let m = HDF5Memory::open(&path).unwrap();
+                let d = t.elapsed();
+                drop(m);
+                d
+            })
+            .collect();
+        opens.sort();
+        let mut mem = HDF5Memory::open(&path).unwrap();
+
+        // Vector-only search: empty text, all weight on the vector stage.
+        let results: Vec<Vec<usize>> = data
+            .queries
+            .iter()
+            .map(|q| {
+                mem.hybrid_search(q, "", 1.0, 0.0, K)
+                    .iter()
+                    .map(|r| r.index)
+                    .collect()
+            })
+            .collect();
+        let hits: usize = results
+            .iter()
+            .zip(&exact)
+            .map(|(got, want)| got.iter().filter(|i| want.contains(i)).count())
+            .sum();
+        let recall = hits as f64 / (K * data.queries.len()) as f64;
+
+        let latency = summarize(
+            (0..N_QUERIES)
+                .map(|i| {
+                    let t = Instant::now();
+                    std::hint::black_box(mem.hybrid_search(
+                        &data.queries[i],
+                        &query_texts[i],
+                        0.4,
+                        0.6,
+                        K,
+                    ));
+                    t.elapsed()
+                })
+                .collect(),
+        );
+        let overlap = match per_variant.first() {
+            Some((_, other)) => {
+                let same: usize = results
+                    .iter()
+                    .zip(other)
+                    .map(|(a, b)| a.iter().filter(|i| b.contains(i)).count())
+                    .sum();
+                format!("{:.4}", same as f64 / (K * data.queries.len()) as f64)
+            }
+            None => "—".into(),
+        };
+        println!(
+            "| {n} | {} | {:.1} | {:.0} | {:.1} | {recall:.4} | {overlap} | {:.3} |",
+            if float16 { "float16" } else { "f32" },
+            mib(file_bytes),
+            millis(checkpoint),
+            millis(opens[1]),
+            millis(latency.p50),
+        );
+        per_variant.push((float16, results));
+    }
+}
+
 // ---------------------------------------------------------------------------
 // Fusion study: does capping the keyword candidate pool change the ranking?
 // ---------------------------------------------------------------------------
@@ -394,11 +902,12 @@ fn fusion_study(n: usize) {
         .map(|(i, c)| text_for(*c, i, &mut rng))
         .collect();
     let bm25 = BM25Index::build(&texts, &vec![0u8; n]);
-    let index = HnswIndex::build_with_metric(
+    let index = HnswIndex::build_with(
         &data.vectors,
         HNSW_M,
         HNSW_EF_CONSTRUCTION,
         DistanceMetric::Cosine,
+        storage(),
     );
 
     let vec_pool = (K * 8).max(64);
@@ -450,6 +959,69 @@ fn fusion_study(n: usize) {
     }
 }
 
+/// What an in-memory store costs, stage by stage. The vectors are the floor:
+/// everything above it is bookkeeping that could in principle be shared.
+fn bench_footprint(n: usize) {
+    let data = make_dataset(n, 0xF007 ^ n as u64);
+    let mut rng = Rng(11);
+    let dir = tempfile::TempDir::new().unwrap();
+    let path = dir.path().join("footprint.h5");
+
+    let base = heap_bytes();
+    let entries: Vec<MemoryEntry> = data
+        .vectors
+        .iter()
+        .enumerate()
+        .map(|(i, v)| MemoryEntry {
+            chunk: text_for(data.cluster_of[i], i, &mut rng),
+            embedding: v.clone(),
+            source_channel: "bench".into(),
+            timestamp: i as f64,
+            session_id: format!("s{}", i % 50),
+            tags: format!("t{i}"),
+        })
+        .collect();
+    let after_entries = heap_bytes();
+
+    let mut config = MemoryConfig::new(path, "bench", DIM);
+    config.quantized_index = INT8.load(std::sync::atomic::Ordering::Relaxed);
+    let mut mem = HDF5Memory::create(config).unwrap();
+    mem.save_batch(entries).unwrap();
+    let after_store = heap_bytes();
+
+    // First query builds the vector and keyword indexes.
+    std::hint::black_box(mem.hybrid_search(&data.queries[0], "record", 0.7, 0.3, K));
+    let after_indexes = heap_bytes();
+
+    // Reopening is the figure that matters for a long-lived process, and the
+    // only one RSS reports honestly: memory freed when the ingest buffers went
+    // away stays in the allocator's pool, so the stage deltas above understate
+    // what was given back.
+    let path = mem.config().path.clone();
+    drop(mem);
+    let before_open = heap_bytes();
+    reset_peak();
+    let reopened = HDF5Memory::open(&path).unwrap();
+    let after_open = heap_bytes();
+    let loaded = after_open.saturating_sub(before_open);
+    // Peak over the open, not just what it leaves behind: a buffer allocated
+    // and freed during the parse never shows up in the live total.
+    let peak = peak_bytes().saturating_sub(before_open);
+    drop(reopened);
+
+    let raw = (n * DIM * 4) as u64;
+    println!(
+        "| {n} | {:.0} | {:.0} | {:.0} | {:.0} | {:.0} | {:.0} | {:.2}x |",
+        mib(raw),
+        mib(after_entries.saturating_sub(base)),
+        mib(after_store.saturating_sub(after_entries)),
+        mib(after_indexes.saturating_sub(after_store)),
+        mib(loaded),
+        mib(peak),
+        loaded as f64 / raw as f64,
+    );
+}
+
 fn main() {
     let args: Vec<String> = std::env::args().skip(1).collect();
     let full = args.iter().any(|a| a == "--full");
@@ -464,6 +1036,60 @@ fn main() {
         }
         return;
     }
+    if args.iter().any(|a| a == "--signing-study") {
+        println!("## Signed checkpoints ({DIM}-dim, float16, int8 index)\n");
+        println!(
+            "| N | checkpoint ms, unsigned | checkpoint ms, signed | signing adds ms | verify ms | file MiB added |"
+        );
+        println!("|---:|---:|---:|---:|---:|---:|");
+        for &n in if full {
+            &[1_000, 10_000, 100_000][..]
+        } else {
+            &[1_000, 10_000][..]
+        } {
+            signing_study(n);
+        }
+        return;
+    }
+    if args.iter().any(|a| a == "--options-study") {
+        println!("## Search options ({DIM}-dim, k = {K}, Hebbian boost off)\n");
+        println!("| N | options | filtered recall@10 | p50 ms | p99 ms |");
+        println!("|---:|---|---:|---:|---:|");
+        for &n in if full {
+            &[10_000, 100_000][..]
+        } else {
+            &[10_000][..]
+        } {
+            options_study(n);
+        }
+        return;
+    }
+    if args.iter().any(|a| a == "--f16-first") {
+        F16_FIRST.store(true, std::sync::atomic::Ordering::Relaxed);
+    }
+    if args.iter().any(|a| a == "--float16-study") {
+        println!("## float16 embedding storage ({DIM}-dim, int8 index, Hebbian boost off)\n");
+        println!(
+            "| N | embeddings | file MiB | checkpoint ms | open ms | recall@10 | top-10 overlap with the other | hybrid p50 ms |"
+        );
+        println!("|---:|---|---:|---:|---:|---:|---:|---:|");
+        for &n in if full {
+            &[1_000, 10_000, 100_000][..]
+        } else {
+            &[1_000, 10_000][..]
+        } {
+            float16_study(n);
+        }
+        return;
+    }
+    if args.iter().any(|a| a == "--int8") {
+        INT8.store(true, std::sync::atomic::Ordering::Relaxed);
+        println!("(int8-quantised index vectors)");
+    }
+    if args.iter().any(|a| a == "--rerank") {
+        RERANK.store(true, std::sync::atomic::Ordering::Relaxed);
+        println!("(candidates re-scored against exact vectors)");
+    }
     if args.iter().any(|a| a == "--uniform") {
         UNIFORM.store(true, std::sync::atomic::Ordering::Relaxed);
         println!("(uniform random data)");
@@ -485,6 +1111,18 @@ fn main() {
 
     let mut json = Vec::new();
     println!("## Search harness");
+
+    if args.iter().any(|a| a == "--footprint") {
+        println!("\n### Resident memory, {DIM}-dim f32\n");
+        println!(
+            "| N | vectors (raw) | entries MiB | store MiB | indexes MiB | reopened MiB | peak during open MiB | reopened / raw |"
+        );
+        println!("|---:|---:|---:|---:|---:|---:|---:|---:|");
+        for &n in sizes {
+            bench_footprint(n);
+        }
+        return;
+    }
     // `--e2e-only` skips the index benchmarks, so the end-to-end section runs
     // in a process that has not already spun up a thread pool.
     if !args.iter().any(|a| a == "--e2e-only") {

crates/clawhdf5-bench/tests/concurrent_read_smoke.rs

@@ -0,0 +1,148 @@
+//! Keeps the concurrent-read harnesses working: runs `concurrent_read`, the
+//! h5py script (threads and processes) and the comparison script end to end
+//! on tiny files. h5py reading the files also checks, element by element at
+//! spot positions, that both harnesses generate the same data and slabs.
+//!
+//! The h5py half is skipped when python3 with h5py is unavailable, unless
+//! `CLAWHDF5_REQUIRE_INTEROP=1`; `CLAWHDF5_PYTHON` picks the interpreter.
+
+use std::path::{Path, PathBuf};
+use std::process::Command;
+
+fn python() -> String {
+    std::env::var("CLAWHDF5_PYTHON").unwrap_or_else(|_| "python3".to_string())
+}
+
+fn interop_required() -> bool {
+    std::env::var("CLAWHDF5_REQUIRE_INTEROP").is_ok_and(|v| v == "1")
+}
+
+fn python_available() -> bool {
+    Command::new(python())
+        .args(["-c", "import h5py, numpy"])
+        .output()
+        .map(|o| o.status.success())
+        .unwrap_or(false)
+}
+
+fn scripts() -> PathBuf {
+    Path::new(env!("CARGO_MANIFEST_DIR")).join("scripts")
+}
+
+fn run(cmd: &mut Command) -> String {
+    let out = cmd.output().expect("spawn");
+    assert!(
+        out.status.success(),
+        "{cmd:?} failed\nSTDOUT:\n{}\nSTDERR:\n{}",
+        String::from_utf8_lossy(&out.stdout),
+        String::from_utf8_lossy(&out.stderr)
+    );
+    String::from_utf8_lossy(&out.stdout).into_owned()
+}
+
+const SMALL: [&str; 8] = [
+    "--threads",
+    "1,2",
+    "--slabs",
+    "8",
+    "--reps",
+    "1",
+    "--slab",
+    "64",
+];
+
+fn results(path: &Path) -> serde_json::Value {
+    serde_json::from_str(&std::fs::read_to_string(path).unwrap()).unwrap()
+}
+
+#[test]
+fn harnesses_run_end_to_end_on_tiny_files() {
+    let dir = tempfile::TempDir::new().unwrap();
+    let data = dir.path().join("data");
+    let claw = dir.path().join("claw.json");
+
+    let bin = env!("CARGO_BIN_EXE_concurrent_read");
+    run(Command::new(bin)
+        .arg("--dir")
+        .arg(&data)
+        .args(["--datasets", "3", "--mib", "1"])
+        .args(SMALL)
+        .arg("--json")
+        .arg(&claw));
+    // Second run reuses the files (and exercises --cold).
+    let out = Command::new(bin)
+        .arg("--dir")
+        .arg(&data)
+        .args(["--datasets", "3", "--mib", "1", "--cold"])
+        .args(SMALL)
+        .output()
+        .unwrap();
+    assert!(out.status.success());
+    assert!(String::from_utf8_lossy(&out.stderr).contains("reusing"));
+
+    let doc = results(&claw);
+    assert_eq!(doc["tool"], "clawhdf5");
+    // 2 layouts x 2 modes x 2 thread counts.
+    assert_eq!(doc["results"].as_array().unwrap().len(), 8);
+    for r in doc["results"].as_array().unwrap() {
+        assert!(r["mb_s"].as_f64().unwrap() > 0.0, "{r}");
+    }
+
+    if !python_available() {
+        assert!(
+            !interop_required(),
+            "CLAWHDF5_REQUIRE_INTEROP=1 but {} has no h5py",
+            python()
+        );
+        eprintln!("skipping the h5py half: no h5py in {}", python());
+        return;
+    }
+    let mut jsons = vec![claw];
+    for executor in ["threads", "processes"] {
+        let out = dir.path().join(format!("h5py-{executor}.json"));
+        run(Command::new(python())
+            .arg(scripts().join("concurrent_read_h5py.py"))
+            .arg("--dir")
+            .arg(&data)
+            .args(["--executor", executor])
+            .args(SMALL)
+            .arg("--json")
+            .arg(&out));
+        let doc = results(&out);
+        assert_eq!(doc["tool"], format!("h5py-{executor}"));
+        assert_eq!(doc["results"].as_array().unwrap().len(), 8);
+        jsons.push(out);
+    }
+    let table = run(Command::new(python())
+        .arg(scripts().join("compare_concurrent_read.py"))
+        .args(&jsons));
+    assert!(table.contains("| deflate | same | 2 |"), "{table}");
+    assert!(table.contains("clawhdf5 / h5py-processes"), "{table}");
+
+    // A different workload must not be compared.
+    let other = dir.path().join("other.json");
+    run(Command::new(python())
+        .arg(scripts().join("concurrent_read_h5py.py"))
+        .arg("--dir")
+        .arg(&data)
+        .args([
+            "--threads",
+            "1",
+            "--slabs",
+            "4",
+            "--reps",
+            "1",
+            "--slab",
+            "64",
+        ])
+        .arg("--json")
+        .arg(&other));
+    let out = Command::new(python())
+        .arg(scripts().join("compare_concurrent_read.py"))
+        .arg(&jsons[0])
+        .arg(&other)
+        .output()
+        .unwrap();
+    assert!(!out.status.success());
+    assert!(String::from_utf8_lossy(&out.stderr).contains("slabs"));
+}

crates/clawhdf5-cli/Cargo.toml

@@ -1,7 +1,8 @@
 [package]
 name = "clawhdf5-cli"
-version = "2.5.0"
+version = "2.7.0"
 edition = "2024"
+rust-version.workspace = true
 license = "MIT"
 description = "CLI for clawhdf5 agent memory — create, save, search, recall, stats"
 repository = "https://git.redclaw.dev/quantumclaw/clawhdf5"
@@ -14,7 +15,7 @@ name = "clawhdf5"
 path = "src/main.rs"
 
 [dependencies]
-clawhdf5-agent = { path = "../clawhdf5-agent", version = "2.5.0" }
+clawhdf5-agent = { path = "../clawhdf5-agent", version = "2.7.0" }
 clap = { version = "4", features = ["derive", "env"] }
 serde_json = "1"
 serde = { workspace = true }

crates/clawhdf5-cli/README.md

@@ -0,0 +1,49 @@
+# clawhdf5-cli
+
+The `clawhdf5` command: create, fill, search and inspect a
+[`clawhdf5-agent`](../clawhdf5-agent/README.md) memory store from the
+shell. Output is JSON. (For general HDF5 files use `h5rs` from
+[`clawhdf5-tools`](../clawhdf5-tools/README.md).)
+
+```bash
+cargo install --path crates/clawhdf5-cli     # installs `clawhdf5`; not on crates.io yet
+# or: cargo run -p clawhdf5-cli -- --help
+```
+
+No C is compiled.
+
+## Commands
+
+The store is `--path FILE` (or `CLAWHDF5_PATH`) before the subcommand.
+
+| Command | What |
+|---|---|
+| `create [--agent-id ID] [--dim N] [--wal] [--f32] [--f32-index]` | a new store (dimension 384 by default); float16 embeddings and an int8 index copy unless `--f32` / `--f32-index`. The WAL is off unless `--wal` (the library's default is on), so each save is checkpointed at once |
+| `save [--json '{...}']` | save one entry, from `--json` or stdin: `{"chunk", "embedding", "source_channel", "timestamp", "session_id", "tags"}` |
+| `search --embedding '[...]' [--query TEXT] [-k N] [--vector-weight W] [--keyword-weight W]` | hybrid search (defaults 5 results, weights 0.7 / 0.3) |
+| `recall INDEX` | one entry by index |
+| `stats` | counts and configuration |
+| `flush-wal` | checkpoint the WAL into the `.h5` |
+| `agents-md [--output FILE]` | generate an `AGENTS.md` from the store |
+| `export` | every entry as JSON lines |
+| `snapshot DEST` | a copy of the store's `.h5` file |
+| `keygen --out FILE` | a new Ed25519 signing key (64 hex characters, created owner-only on Unix) |
+| `verify --public-key HEX_OR_FILE` | check a signed store; exit status 2 if it does not verify |
+
+`recall`, `stats`, `agents-md` and `export` open the store read-only
+(no lock, nothing written), so they work while another process has it
+open. `save`, `search` (which records activation boosts) and `flush-wal`
+open it for writing and take the store's lock. With
+`--signing-key FILE` (or `CLAWHDF5_SIGNING_KEY`) every checkpoint a command
+makes is signed; a signed store refuses to checkpoint without the key.
+
+```bash
+clawhdf5 --path mem.h5 create --agent-id demo --dim 3
+echo '{"chunk":"hello","embedding":[0.1,0.2,0.3],"source_channel":"cli","timestamp":0,"session_id":"s1","tags":""}' \
+  | clawhdf5 --path mem.h5 save
+clawhdf5 --path mem.h5 search --embedding '[0.1,0.2,0.3]' --query hello -k 3
+```
+
+## License
+
+MIT

crates/clawhdf5-cli/src/main.rs

@@ -1,15 +1,22 @@
-use std::path::PathBuf;
+use std::path::{Path, PathBuf};
 
 use clap::{Parser, Subcommand};
+use clawhdf5_agent::signing::{self, SigningKey, VerifyingKey};
 use clawhdf5_agent::{AgentMemory, HDF5Memory, MemoryConfig, MemoryEntry};
 
 /// ClawhDF5 — HDF5-backed cognitive memory for AI agents
 #[derive(Parser)]
 #[command(name = "clawhdf5", version, about)]
 struct Cli {
-    /// Path to the .h5 memory file
+    /// Path to the .h5 memory file (not needed for `keygen`)
     #[arg(short, long, env = "CLAWHDF5_PATH")]
-    path: PathBuf,
+    path: Option<PathBuf>,
+
+    /// File holding an Ed25519 signing key (64 hex characters, from
+    /// `keygen`). Every checkpoint this command makes is then signed; a
+    /// signed store refuses to checkpoint without it.
+    #[arg(long, env = "CLAWHDF5_SIGNING_KEY", global = true)]
+    signing_key: Option<PathBuf>,
 
     #[command(subcommand)]
     command: Commands,
@@ -28,6 +35,22 @@ enum Commands {
         /// Enable write-ahead log
         #[arg(long)]
         wal: bool,
+        /// Hold the vector index's copy of the embeddings as f32 instead of
+        /// the default int8 (which uses a quarter of the memory and is faster
+        /// at equal recall)
+        #[arg(long)]
+        f32_index: bool,
+        /// Accepted for compatibility; int8 is now the default
+        #[arg(long, hide = true, conflicts_with = "f32_index")]
+        quantized_index: bool,
+        /// Store embeddings as full-precision f32 instead of the default
+        /// half precision (float16: half the bytes, about three significant
+        /// digits, values within ±65504)
+        #[arg(long)]
+        f32: bool,
+        /// Accepted for compatibility; float16 is now the default
+        #[arg(long, hide = true, conflicts_with = "f32")]
+        float16: bool,
     },
     /// Save a memory entry (reads JSON from stdin or --json)
     Save {
@@ -75,6 +98,38 @@ enum Commands {
         /// Destination path
         dest: PathBuf,
     },
+    /// Generate an Ed25519 signing key for signed checkpoints
+    Keygen {
+        /// Where to write the secret key (created new, owner-only on Unix)
+        #[arg(long)]
+        out: PathBuf,
+    },
+    /// Verify a signed store against a public key; exit status 2 if not valid
+    Verify {
+        /// The trusted public key: 64 hex characters, or a file holding them
+        #[arg(long)]
+        public_key: String,
+    },
+}
+
+fn read_signing_key(path: &Path) -> Result<SigningKey, Box<dyn std::error::Error>> {
+    let text = std::fs::read_to_string(path)
+        .map_err(|e| format!("cannot read signing key {}: {e}", path.display()))?;
+    let bytes = signing::from_hex::<32>(&text)
+        .ok_or_else(|| format!("{} is not a 64-hex-character key", path.display()))?;
+    Ok(SigningKey::from_bytes(&bytes))
+}
+
+/// Open for writing, with the signing key applied if one was given.
+fn open_writable(
+    path: &Path,
+    key: &Option<SigningKey>,
+) -> Result<HDF5Memory, Box<dyn std::error::Error>> {
+    let mut mem = HDF5Memory::open(path)?;
+    if let Some(k) = key {
+        mem.set_signing_key(k.clone());
+    }
+    Ok(mem)
 }
 
 fn main() {
@@ -87,17 +142,76 @@ fn main() {
 }
 
 fn run(cli: Cli) -> Result<(), Box<dyn std::error::Error>> {
+    if let Commands::Keygen { out } = &cli.command {
+        let key = signing::generate_key();
+        let mut opts = std::fs::OpenOptions::new();
+        opts.write(true).create_new(true);
+        #[cfg(unix)]
+        {
+            use std::os::unix::fs::OpenOptionsExt;
+            opts.mode(0o600);
+        }
+        use std::io::Write;
+        let mut f = opts
+            .open(out)
+            .map_err(|e| format!("cannot create {}: {e}", out.display()))?;
+        writeln!(f, "{}", signing::to_hex(&key.to_bytes()))?;
+        let j = serde_json::json!({
+            "status": "generated",
+            "secret_key_file": out.display().to_string(),
+            "public_key": signing::to_hex(&key.verifying_key().to_bytes()),
+        });
+        println!("{}", serde_json::to_string_pretty(&j)?);
+        return Ok(());
+    }
+    let path = cli
+        .path
+        .clone()
+        .ok_or("--path (or CLAWHDF5_PATH) is required")?;
+    let key = cli
+        .signing_key
+        .as_deref()
+        .map(read_signing_key)
+        .transpose()?;
     match cli.command {
-        Commands::Create { agent_id, dim, wal } => {
-            let mut config = MemoryConfig::new(cli.path.clone(), &agent_id, dim);
+        Commands::Create {
+            agent_id,
+            dim,
+            wal,
+            f32_index,
+            quantized_index: _,
+            f32,
+            float16: _,
+        } => {
+            let mut config = MemoryConfig::new(path.clone(), &agent_id, dim);
             config.wal_enabled = wal;
-            let mem = HDF5Memory::create(config)?;
+            // As with --f32-index: only ever switch the library default off.
+            if f32 {
+                config.float16 = false;
+            }
+            let config_float16 = config.float16;
+            // Only ever switch *off* the library default: assigning the flag
+            // outright would force every CLI-created store back to f32 unless
+            // the caller knew to ask for int8.
+            if f32_index {
+                config.quantized_index = false;
+            }
+            let config_quantized = config.quantized_index;
+            let mut mem = HDF5Memory::create(config)?;
+            // Sign straight away, so the store is never on disk unsigned.
+            if let Some(k) = &key {
+                mem.set_signing_key(k.clone());
+                mem.flush_wal()?;
+            }
             let j = serde_json::json!({
                 "status": "created",
-                "path": cli.path.display().to_string(),
+                "path": path.display().to_string(),
                 "agent_id": agent_id,
                 "embedding_dim": dim,
                 "wal_enabled": wal,
+                "quantized_index": config_quantized,
+                "float16": config_float16,
+                "signed": mem.is_signed(),
                 "count": mem.count(),
             });
             println!("{}", serde_json::to_string_pretty(&j)?);
@@ -114,7 +228,7 @@ fn run(cli: Cli) -> Result<(), Box<dyn std::error::Error>> {
                 }
             };
             let entry: MemoryEntry = serde_json::from_str(&input)?;
-            let mut mem = HDF5Memory::open(&cli.path)?;
+            let mut mem = open_writable(&path, &key)?;
             let idx = mem.save(entry)?;
             let j = serde_json::json!({ "status": "saved", "index": idx, "count": mem.count() });
             println!("{}", serde_json::to_string(&j)?);
@@ -128,7 +242,7 @@ fn run(cli: Cli) -> Result<(), Box<dyn std::error::Error>> {
             keyword_weight,
         } => {
             let emb: Vec<f32> = serde_json::from_str(&embedding)?;
-            let mut mem = HDF5Memory::open(&cli.path)?;
+            let mut mem = open_writable(&path, &key)?;
             let results = mem.hybrid_search(&emb, &query, vector_weight, keyword_weight, top_k);
             let j: Vec<serde_json::Value> = results
                 .iter()
@@ -146,7 +260,7 @@ fn run(cli: Cli) -> Result<(), Box<dyn std::error::Error>> {
         }
 
         Commands::Recall { index } => {
-            let mem = HDF5Memory::open_read_only(&cli.path)?;
+            let mem = HDF5Memory::open_read_only(&path)?;
             match mem.get_chunk(index) {
                 Some(content) => {
                     let j = serde_json::json!({ "index": index, "chunk": content });
@@ -160,22 +274,23 @@ fn run(cli: Cli) -> Result<(), Box<dyn std::error::Error>> {
         }
 
         Commands::Stats => {
-            let mem = HDF5Memory::open_read_only(&cli.path)?;
+            let mem = HDF5Memory::open_read_only(&path)?;
             let cfg = mem.config();
             let j = serde_json::json!({
-                "path": cli.path.display().to_string(),
+                "path": path.display().to_string(),
                 "agent_id": cfg.agent_id,
                 "embedding_dim": cfg.embedding_dim,
                 "count": mem.count(),
                 "active": mem.count_active(),
                 "wal_enabled": cfg.wal_enabled,
                 "wal_pending": mem.wal_pending_count(),
+                "signed": mem.is_signed(),
             });
             println!("{}", serde_json::to_string_pretty(&j)?);
         }
 
         Commands::FlushWal => {
-            let mut mem = HDF5Memory::open(&cli.path)?;
+            let mut mem = open_writable(&path, &key)?;
             let before = mem.wal_pending_count();
             mem.flush_wal()?;
             let j = serde_json::json!({
@@ -187,7 +302,7 @@ fn run(cli: Cli) -> Result<(), Box<dyn std::error::Error>> {
         }
 
         Commands::AgentsMd { output } => {
-            let mem = HDF5Memory::open_read_only(&cli.path)?;
+            let mem = HDF5Memory::open_read_only(&path)?;
             let md = mem.generate_agents_md();
             match output {
                 Some(p) => {
@@ -199,7 +314,7 @@ fn run(cli: Cli) -> Result<(), Box<dyn std::error::Error>> {
         }
 
         Commands::Export => {
-            let mem = HDF5Memory::open_read_only(&cli.path)?;
+            let mem = HDF5Memory::open_read_only(&path)?;
             for i in 0..mem.count() {
                 if let Some(chunk) = mem.get_chunk(i) {
                     let j = serde_json::json!({ "index": i, "chunk": chunk });
@@ -208,11 +323,44 @@ fn run(cli: Cli) -> Result<(), Box<dyn std::error::Error>> {
             }
         }
 
+        Commands::Keygen { .. } => unreachable!("handled before opening a store"),
+
+        Commands::Verify { public_key } => {
+            let text = if Path::new(&public_key).is_file() {
+                std::fs::read_to_string(&public_key)?
+            } else {
+                public_key
+            };
+            let bytes = signing::from_hex::<32>(&text)
+                .ok_or("--public-key must be 64 hex characters or a file holding them")?;
+            let trusted = VerifyingKey::from_bytes(&bytes)?;
+            let r = HDF5Memory::verify(&path, &trusted)?;
+            let j = serde_json::json!({
+                "valid": r.is_valid(),
+                "signed": r.signed,
+                "key_matches": r.key_matches,
+                "signature_valid": r.signature_valid,
+                "records_match": r.records_match,
+                "settings_match": r.settings_match,
+                "sessions_match": r.sessions_match,
+                "graph_match": r.graph_match,
+                "changed_records": r.changed_records,
+                "record_count": r.record_count,
+                "signed_record_count": r.signed_record_count,
+                "signed_by": r.public_key.map(|k| signing::to_hex(&k)),
+                "wal_entries_unsigned": r.wal_entries_unsigned,
+            });
+            println!("{}", serde_json::to_string_pretty(&j)?);
+            if !r.is_valid() {
+                std::process::exit(2);
+            }
+        }
+
         Commands::Snapshot { dest } => {
-            let _result = clawhdf5_agent::storage::snapshot_file(&cli.path, &dest)?;
+            let _result = clawhdf5_agent::storage::snapshot_file(&path, &dest)?;
             let j = serde_json::json!({
                 "status": "snapshot_created",
-                "source": cli.path.display().to_string(),
+                "source": path.display().to_string(),
                 "dest": dest.display().to_string(),
             });
             println!("{}", serde_json::to_string(&j)?);

crates/clawhdf5-derive/Cargo.toml

@@ -1,8 +1,9 @@
 [package]
 name = "clawhdf5-derive"
-version = "2.5.0"
+version = "2.7.0"
 edition = "2024"
-description = "Derive macros for rustyhdf5 HDF5 traits"
+rust-version.workspace = true
+description = "Derive macro (H5Type) for clawhdf5 compound types"
 license = "MIT"
 repository = "https://git.redclaw.dev/quantumclaw/clawhdf5"
 readme = "README.md"

crates/clawhdf5-derive/README.md

@@ -1,28 +1,50 @@
 # clawhdf5-derive
 
-[![crates.io](https://img.shields.io/crates/v/clawhdf5-derive.svg)](https://crates.io/crates/clawhdf5-derive)
-[![docs.rs](https://docs.rs/clawhdf5-derive/badge.svg)](https://docs.rs/clawhdf5-derive)
+`#[derive(H5Type)]`: maps a Rust struct with named fields to an HDF5
+compound datatype. The derive generates three inherent methods:
 
-Derive macros for clawhdf5 HDF5 traits.
+- `hdf5_datatype() -> clawhdf5_format::datatype::Datatype` — the
+  `Datatype::Compound` (members in field order, packed, little-endian);
+- `to_bytes(&self) -> Vec<u8>` — one element in that layout;
+- `from_bytes(&[u8]) -> Self` — the reverse (panics if the slice is shorter
+  than the compound).
 
-## Features
+Supported field types: `f32`, `f64`, `i8`–`i64`, `u8`–`u64`, `bool`
+(stored as `u8`) and fixed-size arrays `[T; N]` of those numeric types.
+Tuple structs, enums and nested structs are refused at compile time.
 
-- `#[derive(HDF5Type)]` for automatic HDF5 datatype mapping
-- Struct-to-compound-type derivation
+The generated code names `clawhdf5_format`, so the crate using the derive
+must depend on [`clawhdf5-format`](../clawhdf5-format/README.md) too. Not
+on crates.io yet:
 
-## Usage
+```toml
+[dependencies]
+clawhdf5-derive = { git = "https://git.redclaw.dev/quantumclaw/clawhdf5" }
+clawhdf5-format = { git = "https://git.redclaw.dev/quantumclaw/clawhdf5" }
+```
+
+## Example
 
 ```rust
-use clawhdf5_derive::HDF5Type;
+use clawhdf5_derive::H5Type;
+use clawhdf5_format::datatype::Datatype;
 
-#[derive(HDF5Type)]
+#[derive(H5Type, Debug, PartialEq)]
 struct Point {
-    x: f64,
-    y: f64,
-    z: f64,
+    id: u32,
+    pos: [f64; 3],
+    valid: bool,
 }
+
+let p = Point { id: 7, pos: [1.0, 2.0, 3.0], valid: true };
+let bytes = p.to_bytes();
+assert_eq!(bytes.len(), 4 + 24 + 1);
+assert_eq!(Point::from_bytes(&bytes), p);
+assert!(matches!(Point::hdf5_datatype(), Datatype::Compound { size: 29, .. }));
 ```
 
+Tests: `crates/clawhdf5-format/tests/derive_tests.rs`.
+
 ## License
 
 MIT

crates/clawhdf5-filters/Cargo.toml

@@ -1,7 +1,8 @@
 [package]
 name = "clawhdf5-filters"
-version = "2.5.0"
+version = "2.7.0"
 edition = "2024"
+rust-version.workspace = true
 description = "Filter and compression pipeline for clawhdf5"
 license = "MIT"
 repository = "https://git.redclaw.dev/quantumclaw/clawhdf5"
@@ -25,8 +26,12 @@ name = "compression_bench"
 harness = false
 
 [features]
-default = ["fast-deflate"]
+# Pure-Rust zlib-rs by default; `fast-deflate` (zlib-ng, C) overrides it.
+default = ["zlib-rs"]
 fast-deflate = ["flate2/zlib-ng"]
 system-zlib = ["flate2/zlib-default"]
-zlib-rs = ["flate2/zlib-rs"]
+# `runtime_detection` gives zlib-rs `std`, which it needs to detect and use
+# SIMD at runtime. flate2 enables it by default, but we build flate2 with
+# default-features = false, and without it zlib-rs inflates 3.5x slower.
+zlib-rs = ["flate2/zlib-rs", "flate2/runtime_detection"]
 apple-compression = []

crates/clawhdf5-filters/README.md

@@ -1,24 +1,55 @@
 # clawhdf5-filters
 
-[![crates.io](https://img.shields.io/crates/v/clawhdf5-filters.svg)](https://crates.io/crates/clawhdf5-filters)
-[![docs.rs](https://docs.rs/clawhdf5-filters/badge.svg)](https://docs.rs/clawhdf5-filters)
+Standalone deflate (zlib) compression and decompression with a choice of
+backend: pure-Rust zlib-rs (default), zlib-ng, Apple's Compression
+framework, or miniz_oxide.
 
-Filter and compression pipeline for clawhdf5.
+This crate holds **deflate backends only**. The HDF5 filter pipeline, the
+filter registry and every other codec (shuffle, Fletcher-32, N-Bit,
+scale-offset, LZ4, Zstd, SZIP, pcodec, LZF, bitshuffle, bzip2, Blosc,
+Blosc2, ZFP) live in [`clawhdf5-format`](../clawhdf5-format/README.md),
+which calls flate2 itself and selects its deflate backend with its own
+features. No library crate of the workspace depends on this one (the
+`clawhdf5` facade uses it only in tests).
+
+Not on crates.io yet; depend on it from git:
+
+```toml
+[dependencies]
+clawhdf5-filters = { git = "https://git.redclaw.dev/quantumclaw/clawhdf5" }
+```
+
+## API
+
+```rust
+use clawhdf5_filters::{deflate_backend, deflate_compress, deflate_decompress};
+
+let data: Vec<u8> = (0..10_000u32).map(|i| (i % 251) as u8).collect();
+let compressed = deflate_compress(&data, 6).unwrap();
+// The second argument bounds the output: the expected decompressed size.
+let decompressed = deflate_decompress(&compressed, data.len()).unwrap();
+assert_eq!(decompressed, data);
+println!("backend: {}", deflate_backend()); // "zlib-rs" by default
+```
+
+Also `deflate_compress_miniz`/`deflate_decompress_miniz` (always
+miniz_oxide) and `fast_deflate::{compress, decompress, active_backend}`.
 
 ## Features
 
-- DEFLATE compression/decompression
-- Fast deflate via zlib-ng (`fast-deflate` feature)
-- Apple Compression framework support (`apple-compression` feature)
+Backend priority: `apple-compression` (macOS only) > zlib-ng > zlib-rs >
+miniz_oxide (with none enabled).
 
-## Usage
+| Feature | Default | Backend | Builds C |
+|---|---|---|---|
+| `zlib-rs` | yes | zlib-rs through flate2, with `runtime_detection` (needed for its SIMD) | no |
+| `fast-deflate` | no | zlib-ng through flate2 | yes (cmake) |
+| `system-zlib` | no | the system zlib through flate2 | yes (`libz-sys`) |
+| `apple-compression` | no | Apple Compression framework, macOS only (ignored elsewhere) | no (links a system framework) |
 
-```rust
-use clawhdf5_filters::{deflate_decode, deflate_encode};
-
-let compressed = deflate_encode(&data, 6).unwrap();
-let decompressed = deflate_decode(&compressed).unwrap();
-```
+zlib-rs matches zlib-ng on HDF5 reads and writes and produces
+byte-identical output: see "Deflate backend" in
+[`BENCHMARKS.md`](../../BENCHMARKS.md).
 
 ## License
 

crates/clawhdf5-filters/src/fast_deflate.rs

@@ -1,12 +1,13 @@
-//! Fast deflate backends: Apple Compression Framework and zlib-ng.
+//! Deflate backends: Apple Compression Framework, zlib-ng and zlib-rs.
 //!
 //! Backend selection priority (decompression & compression):
 //! 1. Apple Compression Framework (macOS only, `apple-compression` feature)
-//! 2. flate2 with zlib-ng backend (`fast-deflate` feature) or miniz_oxide (default)
+//! 2. flate2 with zlib-ng (`fast-deflate`), else zlib-rs (`zlib-rs`, the
+//!    default), else miniz_oxide
 //!
 //! The Apple Compression Framework uses hardware-accelerated zlib on Apple Silicon
-//! and is typically the fastest option on macOS. zlib-ng is the fastest portable
-//! option and what C HDF5 uses internally.
+//! and is typically the fastest option on macOS. zlib-rs is a pure-Rust port of
+//! zlib-ng; see `BENCHMARKS.md` for how the two compare.
 
 // ---------------------------------------------------------------------------
 // Apple Compression Framework FFI (macOS only)
@@ -243,65 +244,146 @@ mod apple {
 }
 
 // ---------------------------------------------------------------------------
-// Streaming decompression via flate2 (uses zlib-ng when fast-deflate enabled)
+// One-shot (de)compression via flate2 (whichever backend flate2 was built with)
+//
+// The whole input goes to the codec in one call, into an output buffer sized
+// up front. `flate2::read::ZlibDecoder` / `write::ZlibEncoder` stream through a
+// 32 KiB buffer instead, which cost zlib-rs up to 3.7x against zlib-ng on a
+// 1 MB chunk. clawhdf5-format's deflate filter does the same; see
+// `BENCHMARKS.md`, "Deflate backend".
 // ---------------------------------------------------------------------------
 
-/// Streaming decompress with pre-allocated output buffer.
-///
-/// When the output size is known (typical for HDF5 chunks), this avoids
-/// dynamic reallocation by writing directly into a pre-sized buffer.
+/// Decompress into a buffer pre-sized to `output_size`, the expected
+/// decompressed length (known for HDF5 chunks). Output longer than that is an
+/// error, as is a stream that ends early.
 pub(crate) fn flate2_decompress_preallocated(
     data: &[u8],
     output_size: usize,
 ) -> Result<Vec<u8>, String> {
-    use std::io::Read;
-    let mut decoder = flate2::read::ZlibDecoder::new(data);
-    let mut output = vec![0u8; output_size];
-    let mut total_read = 0;
-
-    loop {
-        match decoder.read(&mut output[total_read..]) {
-            Ok(0) => break,
-            Ok(n) => total_read += n,
-            Err(e) => return Err(e.to_string()),
-        }
-    }
-    output.truncate(total_read);
-    Ok(output)
+    inflate_bounded(data, output_size, output_size)
 }
 
 /// Absolute ceiling on decompressed output when the caller has no size hint,
 /// preventing unbounded allocation from a hostile/corrupted zlib stream.
 const MAX_DECOMPRESS_SIZE: usize = 256 * 1024 * 1024;
 
-/// Streaming decompress with dynamic sizing (when output size is unknown).
-///
-/// Bounded by [`MAX_DECOMPRESS_SIZE`] since there is no chunk-size hint to
-/// validate against here — an unbounded `read_to_end` would let a hostile
-/// zlib stream force arbitrarily large allocation (a "zlib bomb").
+/// Decompress with no size hint, bounded by [`MAX_DECOMPRESS_SIZE`] so a
+/// hostile zlib stream cannot force arbitrarily large allocation (a "zlib
+/// bomb").
 pub(crate) fn flate2_decompress_streaming(data: &[u8]) -> Result<Vec<u8>, String> {
-    use std::io::Read;
-    let decoder = flate2::read::ZlibDecoder::new(data);
-    let mut result = Vec::new();
-    decoder
-        .take(MAX_DECOMPRESS_SIZE as u64 + 1)
-        .read_to_end(&mut result)
-        .map_err(|e| e.to_string())?;
-    if result.len() > MAX_DECOMPRESS_SIZE {
-        return Err(format!(
+    let hint = data.len().saturating_mul(4).min(1 << 20);
+    inflate_bounded(data, hint, MAX_DECOMPRESS_SIZE).map_err(|e| {
+        if e.ends_with("exceeds size limit") {
+            format!(
                 "decompressed output exceeds {} MiB limit",
                 MAX_DECOMPRESS_SIZE / 1024 / 1024
-        ));
+            )
+        } else {
+            e
         }
-    Ok(result)
+    })
 }
 
-/// Compress data using flate2 (zlib-ng when fast-deflate enabled, else miniz_oxide).
+/// Inflate a zlib stream, starting from `size_hint` bytes of output and
+/// failing past `limit`.
+fn inflate_bounded(data: &[u8], size_hint: usize, limit: usize) -> Result<Vec<u8>, String> {
+    use flate2::{Decompress, FlushDecompress, Status};
+
+    // One byte of headroom past the limit distinguishes an over-size stream
+    // from one that legitimately ends exactly at the limit.
+    let max_capacity = limit.saturating_add(1);
+    let mut out = Vec::new();
+    out.try_reserve_exact(size_hint.clamp(1, max_capacity))
+        .map_err(|e| format!("deflate: cannot allocate output: {e}"))?;
+
+    let mut inflater = Decompress::new(true);
+    loop {
+        let (in_before, out_before) = (inflater.total_in(), inflater.total_out());
+        let status = inflater
+            .decompress_vec(
+                &data[in_before as usize..],
+                &mut out,
+                FlushDecompress::Finish,
+            )
+            .map_err(|e| format!("deflate: {e}"))?;
+        if out.len() > limit {
+            return Err("deflate: output exceeds size limit".into());
+        }
+        match status {
+            Status::StreamEnd => return Ok(out),
+            Status::Ok | Status::BufError if out.len() == out.capacity() => {
+                let grow = out.capacity().min(max_capacity - out.capacity()).max(1);
+                out.try_reserve_exact(grow)
+                    .map_err(|e| format!("deflate: cannot allocate output: {e}"))?;
+            }
+            Status::Ok | Status::BufError => {
+                if inflater.total_in() as usize >= data.len()
+                    || (inflater.total_in(), inflater.total_out()) == (in_before, out_before)
+                {
+                    return Err("deflate: truncated stream".into());
+                }
+            }
+        }
+    }
+}
+
+/// Largest compression output reserved at its worst-case size up front.
+const DEFLATE_EXACT_BOUND: usize = 64 << 20;
+
+/// Compress data using flate2 (zlib-ng, zlib-rs or miniz_oxide; see module docs).
 pub(crate) fn flate2_compress(data: &[u8], level: u32) -> Result<Vec<u8>, String> {
-    use std::io::Write;
-    let mut encoder = flate2::write::ZlibEncoder::new(Vec::new(), flate2::Compression::new(level));
-    encoder.write_all(data).map_err(|e| e.to_string())?;
-    encoder.finish().map_err(|e| e.to_string())
+    use flate2::{Compress, Compression, FlushCompress, Status};
+
+    // zlib's compressBound, plus the zlib header and trailer.
+    let bound = data.len() + (data.len() >> 12) + (data.len() >> 14) + (data.len() >> 25) + 13 + 6;
+    // flate2's Rust backends (zlib-rs, miniz_oxide) zero the whole spare
+    // capacity on each call, so a large input's worst-case bound would be
+    // memory held for nothing (4 GiB for a 4 GiB chunk that deflates to a
+    // few MiB): past 64 MiB the output starts at 1/16 of the bound and
+    // doubles as needed.
+    let first = if bound <= DEFLATE_EXACT_BOUND {
+        bound
+    } else {
+        bound / 16
+    };
+    let mut out = Vec::new();
+    out.try_reserve_exact(first)
+        .map_err(|e| format!("deflate: cannot allocate output: {e}"))?;
+
+    let mut deflater = Compress::new(Compression::new(level), true);
+    loop {
+        let (in_before, out_before) = (deflater.total_in(), deflater.total_out());
+        let rest = &data[in_before as usize..];
+        // zlib takes at most u32::MAX input bytes per call, and `Finish`
+        // ends the stream after the bytes it took: input of 4 GiB or more
+        // was cut at 4 GiB - 1. Finish only once the rest fits one call.
+        let flush = if rest.len() > u32::MAX as usize {
+            FlushCompress::None
+        } else {
+            FlushCompress::Finish
+        };
+        let status = deflater
+            .compress_vec(rest, &mut out, flush)
+            .map_err(|e| format!("deflate: {e}"))?;
+        match status {
+            Status::StreamEnd => {
+                if out.capacity() - out.len() > DEFLATE_EXACT_BOUND {
+                    out.shrink_to_fit();
+                }
+                return Ok(out);
+            }
+            // Out of room (the bound makes it unreachable below
+            // `DEFLATE_EXACT_BOUND`): grow rather than fail.
+            Status::Ok | Status::BufError if out.len() == out.capacity() => out
+                .try_reserve(out.capacity().max(4096))
+                .map_err(|e| format!("deflate: cannot allocate output: {e}"))?,
+            Status::Ok | Status::BufError => {
+                if (deflater.total_in(), deflater.total_out()) == (in_before, out_before) {
+                    return Err("deflate: encoder made no progress".into());
+                }
+            }
+        }
+    }
 }
 
 // ---------------------------------------------------------------------------
@@ -312,7 +394,7 @@ pub(crate) fn flate2_compress(data: &[u8], level: u32) -> Result<Vec<u8>, String
 ///
 /// Selection order:
 /// 1. Apple Compression Framework (macOS + `apple-compression` feature)
-/// 2. flate2 (zlib-ng with `fast-deflate`, otherwise miniz_oxide)
+/// 2. flate2 (zlib-ng with `fast-deflate`, else zlib-rs, else miniz_oxide)
 ///
 /// When `output_hint` > 0, pre-allocates the output buffer for zero-copy
 /// decompression (avoids reallocation).
@@ -344,7 +426,7 @@ pub fn decompress(data: &[u8], output_hint: usize) -> Result<Vec<u8>, String> {
 ///
 /// Selection order:
 /// 1. Apple Compression Framework (macOS + `apple-compression` feature)
-/// 2. flate2 (zlib-ng with `fast-deflate`, otherwise miniz_oxide)
+/// 2. flate2 (zlib-ng with `fast-deflate`, else zlib-rs, else miniz_oxide)
 pub fn compress(data: &[u8], level: u32) -> Result<Vec<u8>, String> {
     #[cfg(all(target_os = "macos", feature = "apple-compression"))]
     {
@@ -377,9 +459,19 @@ pub fn active_backend() -> &'static str {
     {
         "zlib-ng"
     }
+    // flate2 prefers a C zlib over zlib-rs when both are enabled.
+    #[cfg(all(
+        not(all(target_os = "macos", feature = "apple-compression")),
+        not(feature = "fast-deflate"),
+        feature = "zlib-rs"
+    ))]
+    {
+        "zlib-rs"
+    }
     #[cfg(not(any(
         all(target_os = "macos", feature = "apple-compression"),
-        feature = "fast-deflate"
+        feature = "fast-deflate",
+        feature = "zlib-rs"
     )))]
     {
         "miniz_oxide"
@@ -436,7 +528,7 @@ mod tests {
     fn backend_name_is_set() {
         let name = active_backend();
         assert!(
-            ["miniz_oxide", "zlib-ng", "apple-compression"].contains(&name),
+            ["miniz_oxide", "zlib-rs", "zlib-ng", "apple-compression"].contains(&name),
             "unexpected backend: {name}"
         );
     }

crates/clawhdf5-filters/src/lib.rs

@@ -2,12 +2,14 @@
 //!
 //! Provides deflate (zlib) decompression/compression with multiple backend options:
 //!
-//! - **Default**: `miniz_oxide` (pure Rust, no C dependencies)
-//! - **`fast-deflate` feature**: `zlib-ng` via flate2 (~2-3x faster, matches C HDF5)
+//! - **Default (`zlib-rs` feature)**: `zlib-rs` via flate2 (pure Rust, no C
+//!   dependencies)
+//! - **`fast-deflate` feature**: `zlib-ng` via flate2 (C, built with cmake)
 //! - **`apple-compression` feature**: Apple Compression Framework on macOS
 //!   (hardware-accelerated on Apple Silicon)
+//! - With none of the above: `miniz_oxide` (pure Rust, slower)
 //!
-//! Backend priority: apple-compression > zlib-ng > miniz_oxide.
+//! Backend priority: apple-compression > zlib-ng > zlib-rs > miniz_oxide.
 
 pub mod fast_deflate;
 
@@ -115,7 +117,7 @@ mod tests {
     fn backend_reports_name() {
         let name = deflate_backend();
         assert!(
-            ["miniz_oxide", "zlib-ng", "apple-compression"].contains(&name),
+            ["miniz_oxide", "zlib-rs", "zlib-ng", "apple-compression"].contains(&name),
             "unexpected backend: {name}"
         );
     }

crates/clawhdf5-format/Cargo.toml

@@ -1,7 +1,8 @@
 [package]
 name = "clawhdf5-format"
-version = "2.5.0"
+version = "2.7.0"
 edition = "2024"
+rust-version.workspace = true
 description = "Pure-Rust HDF5 binary format parsing and writing — no C dependencies"
 license = "MIT"
 repository = "https://git.redclaw.dev/quantumclaw/clawhdf5"
@@ -21,18 +22,33 @@ zstd = { version = "0.13", optional = true }
 blake3 = { version = "1", optional = true }
 libaec-sys = { path = "../libaec-sys", version = "0.1", optional = true }
 pco = { version = "1.0", optional = true }
+# Pure-Rust Zstandard, for the plugin filters that embed zstd (bitshuffle,
+# blosc). The `zstd` feature (filter 32015) links libzstd instead.
+ruzstd = { version = "0.9", optional = true }
+# bzip2 with its default backend, libbz2-rs-sys: a pure-Rust port of
+# libbzip2 (no C is compiled, despite the -sys name).
+bzip2 = { version = "0.6", optional = true }
+snap = { version = "1", optional = true }
+
+[target.'cfg(target_os = "linux")'.dependencies]
+# madvise(MADV_HUGEPAGE) for large read buffers (see src/bulk_alloc.rs).
+libc = { version = "0.2", default-features = false }
 
 [dev-dependencies]
+half = { workspace = true }
 serde_json = "1"
 criterion = { workspace = true }
-clawhdf5-derive = { path = "../clawhdf5-derive", version = "2.5.0" }
+clawhdf5-derive = { path = "../clawhdf5-derive", version = "2.7.0" }
 
 [[bench]]
 name = "bench"
 harness = false
 
 [features]
-default = ["std", "checksum", "deflate", "provenance", "fast-deflate", "system-zlib-decompress"]
+# Deflate backend: `zlib-rs` (pure Rust) by default. `fast-deflate` selects
+# zlib-ng instead (C, built with cmake); flate2 prefers a C zlib whenever one
+# is enabled, so turning it on anywhere in the build overrides the default.
+default = ["std", "checksum", "deflate", "provenance", "zlib-rs", "system-zlib-decompress", "lzf"]
 std = []
 checksum = []
 deflate = ["flate2"]
@@ -42,12 +58,34 @@ fast-checksum = ["crc32fast"]
 fast-deflate = ["flate2/zlib-ng"]
 system-zlib = ["flate2/zlib-default"]
 system-zlib-decompress = []
-zlib-rs = ["flate2/zlib-rs"]
+# `runtime_detection` gives zlib-rs `std`, which it needs to detect and use
+# SIMD at runtime. flate2 enables it by default, but we build flate2 with
+# default-features = false, and without it zlib-rs inflates 3.5x slower.
+zlib-rs = ["flate2/zlib-rs", "flate2/runtime_detection"]
 lz4 = ["lz4_flex"]
 zstd = ["dep:zstd"]
 blake3_hash = ["blake3"]
 szip = ["libaec-sys"]
 pcodec = ["dep:pco"]
+# Plugin filters, pure Rust. LZF (32000) is h5py's built-in compression; it
+# has no dependencies, so it is on by default.
+lzf = []
+# Bitshuffle (32008), with its LZ4 and Zstandard modes.
+bitshuffle = ["lz4_flex", "ruzstd"]
+# bzip2 (307).
+bzip2 = ["dep:bzip2", "std"]
+# Blosc 1 (32001) with its BloscLZ, LZ4, Snappy, Zlib and Zstandard codecs.
+blosc = ["lz4_flex", "ruzstd", "snap", "deflate", "std"]
+# Blosc2 (32026), read-only: frames, B2ND arrays, and the Blosc codecs above.
+blosc2 = ["blosc"]
+# ZFP (32013, H5Z-ZFP), read-only: every mode, for int32, int64, float and
+# double fields of 1 to 4 dimensions.
+zfp = []
+# Every plugin filter above.
+plugin-filters = ["lzf", "bitshuffle", "bzip2", "blosc", "blosc2", "zfp"]
+# Test instrumentation: per-thread counts of heap objects read (see
+# `lookup_stats`), so tests can bound the cost of a name lookup.
+lookup-stats = ["std"]
 
 [[bench]]
 name = "parallel_decompress_bench"

crates/clawhdf5-format/README.md

@@ -1,27 +1,108 @@
 # clawhdf5-format
 
-[![crates.io](https://img.shields.io/crates/v/clawhdf5-format.svg)](https://crates.io/crates/clawhdf5-format)
-[![docs.rs](https://docs.rs/clawhdf5-format/badge.svg)](https://docs.rs/clawhdf5-format)
+The HDF5 file format in pure Rust: parsers and writers for every on-disk
+structure, the filter pipeline and its codecs, and the shared type
+definitions the other crates use. Most users want the
+[`clawhdf5`](../clawhdf5/README.md) facade, which wraps this crate in an
+h5py-like API; use this one directly for low-level access or in `no_std`
+code.
 
-Pure-Rust HDF5 binary format parsing and writing — no C dependencies.
+Not on crates.io yet; depend on it from git:
+
+```toml
+[dependencies]
+clawhdf5-format = { git = "https://git.redclaw.dev/quantumclaw/clawhdf5" }
+```
+
+## What is in it
+
+- **Parsing:** superblock v0–v3 (`superblock`, with the superblock
+  extension and metadata cache images, `superblock_ext`), object headers v1
+  and v2 (`object_header`), every header message the readers use
+  (`datatype`, `dataspace`, `data_layout` v1–v4 including virtual datasets,
+  `fill_value`, `attribute`, `link_message`, `shared_message`, ...), groups
+  old and new (`group_v1` symbol tables with local heaps, `group_v2` with
+  fractal heaps and v2 B-trees), and every chunk index (v1 B-tree, single
+  chunk, implicit, fixed array, extensible array, v2 B-tree).
+- **Reading data:** `data_read` (contiguous, compact, chunked),
+  `partial_read` and `selection` (hyperslabs and points), `vl_data`
+  (variable-length strings and sequences through the global heap),
+  `chunk_cache`.
+- **Storage:** the `storage::Storage` trait (`read_at`, `read_ranges`,
+  `len`, `hint`) that every read path goes through, so a file can be read
+  from memory, a file handle or a remote backend
+  ([`clawhdf5-remote`](../clawhdf5-remote/README.md)).
+- **Writing:** `file_writer::FileWriter` and the builders in
+  `type_builders` (datasets, groups, attributes, compound and enum types,
+  links, virtual datasets, creation-order tracking); chunk indexes and
+  dense-storage B-trees of any size (`chunked_write`, `btree_v2_write`,
+  `ea_writer`, and the version-1 chunk B-tree of `btree_v1_write`). Output
+  is read by h5py and h5dump; `FileWriter::libver_bounds` (`libver`) picks
+  the format: HDF5 1.10 by default, or one HDF5 1.8 reads.
+- **Filters:** `filter_pipeline` and `filter_registry` (look up by ID; other
+  IDs can be registered at run time with `register_filter`). Built in:
+  deflate, shuffle, Fletcher-32, N-Bit, scale-offset; behind features LZ4,
+  Zstd, SZIP (decode), pcodec, and the plugin filters LZF, bitshuffle,
+  bzip2, Blosc 1 (read and write), Blosc2 and ZFP (read only).
+- **Shared pieces:** `float16` (the one IEEE half-precision conversion the
+  workspace uses), `provenance` (SHA-256 dataset hashes), `checksum`
+  (Jenkins lookup3 for v2+ structures).
+
+## Example
+
+```rust
+use clawhdf5_format::file_writer::{AttrValue, FileWriter};
+use clawhdf5_format::{group_v2, object_header, signature, superblock};
+
+// Write a file to memory
+let mut fw = FileWriter::new();
+fw.create_dataset("data")
+    .with_f64_data(&[1.0, 2.0, 3.0])
+    .with_shape(&[3])
+    .set_attr("unit", AttrValue::String("m/s".into()));
+let bytes = fw.finish().unwrap();
+
+// Parse it back: superblock -> path -> object header
+let (_user_block, file) = signature::split_user_block(&bytes).unwrap();
+let sb = superblock::Superblock::parse(file, 0).unwrap();
+let addr = group_v2::resolve_path_any(file, &sb, "data").unwrap();
+let hdr = object_header::ObjectHeader::parse(file, addr as usize, sb.offset_size, sb.length_size)
+    .unwrap();
+assert!(!hdr.messages.is_empty());
+```
 
 ## Features
 
-- Zero-copy superblock, object header, and B-tree parsing
-- Chunked dataset read/write with filter pipelines
-- `no_std` support (disable `std` feature)
-- Optional parallel reads via Rayon
-- SHA-256 provenance tracking
+| Feature | Default | What | Builds C |
+|---|---|---|---|
+| `std` | yes | standard library; without it the crate is `no_std` + `alloc` (CI builds it for `thumbv7em-none-eabihf`) | no |
+| `checksum` | yes | verify Jenkins lookup3 checksums | no |
+| `deflate` | yes | deflate through flate2 | no |
+| `zlib-rs` | yes | flate2's pure-Rust zlib-rs backend, with `runtime_detection` (without it zlib-rs loses SIMD and inflates 3.5x slower) | no |
+| `system-zlib-decompress` | yes | macOS only: inflate with the system libz first, falling back to flate2; no effect elsewhere | no (links the system libz on macOS) |
+| `provenance` | yes | SHA-256 provenance hashes | no |
+| `lzf` | yes | LZF (32000) | no |
+| `parallel` | no | rayon-parallel chunk decoding | no |
+| `fast-checksum` | no | hardware CRC32 through `crc32fast` | no |
+| `lz4` | no | LZ4 (32004) | no |
+| `pcodec` | no | pcodec | no |
+| `bitshuffle`, `bzip2`, `blosc` | no | 32008, 307, 32001, read and write | no |
+| `blosc2`, `zfp` | no | 32026, 32013, read only | no |
+| `plugin-filters` | no | all six plugin filters above | no |
+| `lookup-stats` | no | counters for name-lookup benchmarks | no |
+| `zstd` | no | Zstandard (32015) | yes (libzstd) |
+| `szip` | no | SZIP (4) decoding | links the system libaec (`libaec-dev`) |
+| `fast-deflate` | no | zlib-ng | yes (cmake) |
+| `system-zlib` | no | the system zlib | yes (`libz-sys`) |
+| `blake3_hash` | no | `provenance::blake3_hash` | yes (`cc`) |
 
-## Usage
+## Robustness
 
-```rust
-use clawhdf5_format::Superblock;
-
-let data = std::fs::read("data.h5").unwrap();
-let sb = Superblock::from_bytes(&data).unwrap();
-println!("HDF5 version {}.{}", sb.version_major(), sb.version_minor());
-```
+Every parser is meant to return an error, never panic, on hostile input:
+nine cargo-fuzz targets live in [`fuzz/`](fuzz/README.md), the conformance
+sweep includes the HDF Group's CVE corpus
+([`CONFORMANCE.md`](../../CONFORMANCE.md)), and header checks follow
+libhdf5's. Open gaps are in [`docs/known-issues.md`](../../docs/known-issues.md).
 
 ## License
 

crates/clawhdf5-format/fuzz/.gitignore

@@ -1 +1,4 @@
 target/
+corpus/
+artifacts/
+coverage/

crates/clawhdf5-format/fuzz/README.md

@@ -51,10 +51,18 @@ done
 
 ## CI
 
-These targets are **not** run in CI (`.gitea/workflows/ci.yml`) — cargo-fuzz
-requires nightly and each meaningful run takes minutes, which doesn't fit a
-per-PR gate. Run them manually on a schedule (e.g. before a release, or after
-touching parser code) instead.
+These targets are **not** run by the CI workflows (`.gitea/workflows/ci.yml`)
+— cargo-fuzz requires nightly and each meaningful run takes minutes, which
+doesn't fit a per-PR gate. Run them by hand before a release or after
+touching parser code. `scripts/ci-test.sh` has an opt-in smoke run: with
+`CLAWHDF5_FUZZ_SECONDS=N` it runs every target of this crate and of
+`crates/clawhdf5-agent/fuzz` (the WAL parser) for N seconds each.
+
+Other robustness checks that do run: the nightly conformance sweep reads
+the HDF Group's CVE reproducers and fails on any panic, hang, crash or
+out-of-memory ([`conformance/README.md`](../../../conformance/README.md)),
+and `scripts/h5rs-fuzz.sh` runs every `h5rs` subcommand over them, optionally
+on byte-flipped copies.
 
 ## Reproducing Crashes
 

crates/clawhdf5-format/fuzz/fuzz_targets/fuzz_btree_v2.rs

@@ -1,15 +1,36 @@
 #![no_main]
+use clawhdf5_format::btree_v2::{BTreeV2Header, collect_btree_v2_records};
 use libfuzzer_sys::fuzz_target;
 
 fuzz_target!(|data: &[u8]| {
     for &offset_size in &[4u8, 8] {
         for &length_size in &[4u8, 8] {
-            let _ = clawhdf5_format::btree_v2::BTreeV2Header::parse(
-                data,
-                0,
-                offset_size,
-                length_size,
-            );
+            if let Ok(header) = BTreeV2Header::parse(data, 0, offset_size, length_size) {
+                let _ = collect_btree_v2_records(data, &header, offset_size, length_size);
             }
         }
+    }
+
+    // Parsing a header requires a valid checksum, which random input almost
+    // never has, so the traversal behind it went unfuzzed — and that is where
+    // a node listing itself as its own child overflowed the stack. Take the
+    // header fields straight from the input instead and walk the rest.
+    let Some((fields, file)) = data.split_first_chunk::<20>() else {
+        return;
+    };
+    let header = BTreeV2Header {
+        tree_type: fields[0],
+        node_size: u32::from_le_bytes([fields[1], fields[2], fields[3], fields[4]]),
+        record_size: u16::from_le_bytes([fields[5], fields[6]]),
+        depth: u16::from_le_bytes([fields[7], fields[8]]),
+        root_node_address: u64::from(u32::from_le_bytes([
+            fields[9], fields[10], fields[11], fields[12],
+        ])),
+        num_records_in_root: u16::from_le_bytes([fields[13], fields[14]]),
+        total_records: u64::from(u32::from_le_bytes([
+            fields[15], fields[16], fields[17], fields[18],
+        ])),
+    };
+    let offset_size = if fields[19] & 1 == 0 { 4 } else { 8 };
+    let _ = collect_btree_v2_records(file, &header, offset_size, 8);
 });

crates/clawhdf5-format/src/addr.rs

@@ -0,0 +1,121 @@
+//! File address and length → in-memory index conversion.
+//!
+//! HDF5 addresses and lengths are 64-bit; the file is parsed through a
+//! `&[u8]` indexed by `usize`. On a 64-bit target every `u64` fits, but on a
+//! 32-bit one (`wasm32`, `i686`, `thumbv7em`) an address past `usize::MAX`
+//! used to be truncated by an `as usize` cast — silently pointing at another
+//! part of the file — or to panic. [`to_usize`] is the one conversion the
+//! parsers use instead: such an address is a clean
+//! [`FormatError::Overflow`]. It cannot be inside the data anyway: no slice
+//! is longer than `isize::MAX` bytes.
+
+#[cfg(not(feature = "std"))]
+use alloc::format;
+
+use crate::error::FormatError;
+
+/// A file address, offset or length from the file as a `usize` index.
+///
+/// Fails with [`FormatError::Overflow`] when the value does not fit this
+/// platform's `usize` (only possible on targets narrower than 64 bits).
+#[inline]
+pub fn to_usize(value: u64) -> Result<usize, FormatError> {
+    to_index::<usize>(value)
+}
+
+/// A file address for a [`crate::storage::Storage`] read, checked as
+/// [`to_usize`] checks it: the parsers read through 64-bit offsets, but an
+/// address that could not index an in-memory file on this platform is the
+/// same [`FormatError::Overflow`] the slice parsers gave for it.
+#[inline]
+pub fn checked_addr(value: u64) -> Result<u64, FormatError> {
+    to_usize(value).map(|_| value)
+}
+
+/// [`to_usize`] for an index type of any width. `usize` is 64 bits wide on
+/// the hosts CI tests on, where the error path cannot be reached through
+/// `usize`; tests run the same code with `u32` in its place, as on a 32-bit
+/// target.
+#[inline]
+fn to_index<T: TryFrom<u64>>(value: u64) -> Result<T, FormatError> {
+    T::try_from(value).map_err(|_| too_large(value))
+}
+
+/// A count or offset into an in-memory buffer (a codec's progress counter,
+/// a size the writer computed from data it holds) as a `usize`, saturating
+/// at `usize::MAX` instead of truncating.
+///
+/// For values that are bounded by the length of something in memory, so
+/// always fit; if one ever did not, a saturated index fails its bounds check
+/// or allocation instead of silently addressing the wrong bytes. A value
+/// read from the file uses [`to_usize`].
+#[inline]
+pub fn saturating_usize(value: u64) -> usize {
+    saturating_index(value, usize::MAX)
+}
+
+/// [`saturating_usize`] for an index type of any width, whose largest
+/// value is `max` (see [`to_index`]).
+#[inline]
+fn saturating_index<T: TryFrom<u64>>(value: u64, max: T) -> T {
+    T::try_from(value).unwrap_or(max)
+}
+
+#[cold]
+#[inline(never)]
+fn too_large(value: u64) -> FormatError {
+    FormatError::Overflow(format!(
+        "file address or length {value:#x} exceeds this platform's address space"
+    ))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn values_that_fit_convert_exactly() {
+        assert_eq!(to_usize(0), Ok(0));
+        assert_eq!(to_usize(0x1234), Ok(0x1234));
+        assert_eq!(to_usize(usize::MAX as u64), Ok(usize::MAX));
+    }
+
+    #[test]
+    fn saturating_conversion_never_wraps() {
+        assert_eq!(saturating_usize(0), 0);
+        assert_eq!(saturating_usize(0x1234), 0x1234);
+        assert_eq!(saturating_usize(usize::MAX as u64), usize::MAX);
+        // Past usize::MAX (32-bit targets) or at u64::MAX: saturates.
+        assert_eq!(saturating_usize(u64::MAX), usize::MAX);
+    }
+
+    #[test]
+    fn values_past_usize_max_are_an_error_not_truncated() {
+        // Reachable through `usize` only where it is narrower than u64 (no
+        // such target runs tests in CI), so the same conversion is run with
+        // u32 standing in for a 32-bit usize.
+        let max = u64::from(u32::MAX);
+        assert_eq!(to_index::<u32>(max), Ok(u32::MAX));
+        for past in [max + 1, max + 0x10, 0x1_0000_1234, u64::MAX] {
+            let err = to_index::<u32>(past).unwrap_err();
+            assert!(
+                matches!(err, FormatError::Overflow(_)),
+                "{past:#x}: {err:?}"
+            );
+        }
+        // Where an `as` cast would have wrapped to a small, valid-looking
+        // index, it is not returned.
+        assert_eq!(0x1_0000_1234_u64 as u32, 0x1234);
+        assert!(to_index::<u32>(0x1_0000_1234).is_err());
+
+        assert_eq!(saturating_index(max + 1, u32::MAX), u32::MAX);
+        assert_eq!(saturating_index(0x1_0000_1234, u32::MAX), u32::MAX);
+        assert_eq!(saturating_index(0x1234, u32::MAX), 0x1234);
+
+        // And through `usize` itself, whichever width it has here.
+        match (usize::MAX as u64).checked_add(1) {
+            Some(past) => assert!(matches!(to_usize(past), Err(FormatError::Overflow(_)))),
+            None => assert_eq!(to_usize(u64::MAX), Ok(usize::MAX)),
+        }
+    }
+}

crates/clawhdf5-format/src/attribute.rs

@@ -5,8 +5,10 @@ use alloc::{borrow::Cow, string::String, vec::Vec};
 #[cfg(feature = "std")]
 use std::borrow::Cow;
 
+use crate::addr::to_usize;
 use crate::attribute_info::AttributeInfoMessage;
-use crate::btree_v2::{BTreeV2Header, collect_btree_v2_records};
+use crate::btree_v2::{BTreeV2Header, collect_btree_v2_records_in, find_btree_v2_records_in};
+use crate::checksum::jenkins_lookup3;
 use crate::data_read;
 use crate::dataspace::Dataspace;
 use crate::datatype::Datatype;
@@ -15,6 +17,7 @@ use crate::fractal_heap::FractalHeapHeader;
 use crate::message_type::MessageType;
 use crate::object_header::ObjectHeader;
 use crate::shared_message;
+use crate::storage::Storage;
 use crate::vl_data;
 
 /// A parsed HDF5 attribute message.
@@ -50,7 +53,7 @@ impl AttributeMessage {
     ///
     /// `length_size` is needed for dataspace dimension parsing.
     pub fn parse(data: &[u8], length_size: u8) -> Result<AttributeMessage, FormatError> {
-        Self::parse_impl(data, length_size, None)
+        Self::parse_impl(data, length_size, None::<(&[u8], u8)>)
     }
 
     /// [`AttributeMessage::parse`] with access to the rest of the file, which
@@ -65,13 +68,24 @@ impl AttributeMessage {
         offset_size: u8,
         length_size: u8,
     ) -> Result<AttributeMessage, FormatError> {
-        Self::parse_impl(data, length_size, Some((file_data, offset_size)))
+        Self::parse_in_storage(data, file_data, offset_size, length_size)
     }
 
-    fn parse_impl(
+    /// [`AttributeMessage::parse_in_file`] with the file behind any
+    /// [`Storage`].
+    pub fn parse_in_storage<S: Storage + ?Sized>(
+        data: &[u8],
+        file: &S,
+        offset_size: u8,
+        length_size: u8,
+    ) -> Result<AttributeMessage, FormatError> {
+        Self::parse_impl(data, length_size, Some((file, offset_size)))
+    }
+
+    fn parse_impl<S: Storage + ?Sized>(
         data: &[u8],
         length_size: u8,
-        file: Option<(&[u8], u8)>,
+        file: Option<(&S, u8)>,
     ) -> Result<AttributeMessage, FormatError> {
         ensure_len(data, 0, 2)?;
         let version = data[0];
@@ -86,19 +100,19 @@ impl AttributeMessage {
 
     /// The bytes of an embedded datatype/dataspace message, following the
     /// shared-message reference when `shared` is set.
-    fn embedded_message<'a>(
+    fn embedded_message<'a, S: Storage + ?Sized>(
         bytes: &'a [u8],
         shared: bool,
         msg_type: MessageType,
         length_size: u8,
-        file: Option<(&[u8], u8)>,
+        file: Option<(&S, u8)>,
     ) -> Result<Cow<'a, [u8]>, FormatError> {
         if !shared {
             return Ok(Cow::Borrowed(bytes));
         }
         let (file_data, offset_size) = file.ok_or(FormatError::UnresolvedSharedMessage)?;
-        let shared_ref = shared_message::parse_shared_ref(bytes, offset_size)?;
-        shared_message::resolve_shared_message(
+        let shared_ref = shared_message::parse_shared_ref_sized(bytes, offset_size, length_size)?;
+        shared_message::resolve_shared_message_in(
             file_data,
             &shared_ref,
             msg_type,
@@ -143,10 +157,10 @@ impl AttributeMessage {
         })
     }
 
-    fn parse_v2(
+    fn parse_v2<S: Storage + ?Sized>(
         data: &[u8],
         length_size: u8,
-        file: Option<(&[u8], u8)>,
+        file: Option<(&S, u8)>,
     ) -> Result<AttributeMessage, FormatError> {
         // Flags: bit 0 = datatype is shared, bit 1 = dataspace is shared.
         let flags = data.get(1).copied().unwrap_or(0);
@@ -197,10 +211,10 @@ impl AttributeMessage {
         })
     }
 
-    fn parse_v3(
+    fn parse_v3<S: Storage + ?Sized>(
         data: &[u8],
         length_size: u8,
-        file: Option<(&[u8], u8)>,
+        file: Option<(&S, u8)>,
     ) -> Result<AttributeMessage, FormatError> {
         // Flags: bit 0 = datatype is shared, bit 1 = dataspace is shared.
         let flags = data.get(1).copied().unwrap_or(0);
@@ -322,9 +336,19 @@ impl AttributeMessage {
         file_data: &[u8],
         offset_size: u8,
         length_size: u8,
+    ) -> Result<Vec<String>, FormatError> {
+        self.read_vl_strings_in(file_data, offset_size, length_size)
+    }
+
+    /// [`Self::read_vl_strings`] over any [`Storage`].
+    pub fn read_vl_strings_in<S: Storage + ?Sized>(
+        &self,
+        file_data: &S,
+        offset_size: u8,
+        length_size: u8,
     ) -> Result<Vec<String>, FormatError> {
         let num_elements = self.dataspace.num_elements();
-        vl_data::read_vl_strings(
+        vl_data::read_vl_strings_in(
             file_data,
             &self.raw_data,
             num_elements,
@@ -341,7 +365,8 @@ fn compute_raw_data(
     dataspace: &Dataspace,
     datatype: &Datatype,
 ) -> Vec<u8> {
-    let num_elements = dataspace.num_elements() as usize;
+    // Saturating, like the product: the size is capped at what is there.
+    let num_elements = usize::try_from(dataspace.num_elements()).unwrap_or(usize::MAX);
     let elem_size = datatype.type_size() as usize;
     let expected_size = num_elements.saturating_mul(elem_size);
     let available = data.len().saturating_sub(pos);
@@ -362,6 +387,18 @@ fn extract_name(bytes: &[u8]) -> String {
     String::from_utf8_lossy(&bytes[..end]).into_owned()
 }
 
+/// An attribute's datatype gets libhdf5's extra check for a header without
+/// a checksum (see [`Datatype::check_unused_bits`]).
+fn check_in_header(
+    attr: AttributeMessage,
+    header: &ObjectHeader,
+) -> Result<AttributeMessage, FormatError> {
+    if header.version == 1 {
+        attr.datatype.check_unused_bits()?;
+    }
+    Ok(attr)
+}
+
 /// Extract all attribute messages from an object header.
 pub fn extract_attributes(
     header: &ObjectHeader,
@@ -371,7 +408,7 @@ pub fn extract_attributes(
     for msg in &header.messages {
         if msg.msg_type == MessageType::Attribute {
             let attr = AttributeMessage::parse(&msg.data, length_size)?;
-            attrs.push(attr);
+            attrs.push(check_in_header(attr, header)?);
         }
     }
     Ok(attrs)
@@ -394,57 +431,337 @@ pub fn find_attribute<'a>(
 ///
 /// Use this instead of `extract_attributes` when reading files that may use dense storage
 /// (e.g., objects with many attributes, typically >8).
+///
+/// Fails if any attribute cannot be read; see [`extract_attributes_tolerant`]
+/// to read the others.
 pub fn extract_attributes_full(
     file_data: &[u8],
     header: &ObjectHeader,
     offset_size: u8,
     length_size: u8,
 ) -> Result<Vec<AttributeMessage>, FormatError> {
-    let mut attrs = Vec::new();
+    extract_attributes_full_in(file_data, header, offset_size, length_size)
+}
 
-    // Collect compact attributes (inline in OH)
+/// [`extract_attributes_full`] over any [`Storage`]. Dense attribute
+/// storage is indexed by a v2 B-tree, which is not read over [`Storage`]
+/// yet: on a backend without the whole file in memory an object with dense
+/// attributes is [`FormatError::ContiguousStorageRequired`].
+pub fn extract_attributes_full_in<S: Storage + ?Sized>(
+    file: &S,
+    header: &ObjectHeader,
+    offset_size: u8,
+    length_size: u8,
+) -> Result<Vec<AttributeMessage>, FormatError> {
+    extract_attributes_with(file, header, offset_size, length_size, &mut Err)
+}
+
+/// Like [`extract_attributes_full`], but an attribute that cannot be read
+/// (a corrupt or unsupported attribute message, or a heap object that cannot
+/// be located) is left out and its error returned alongside the attributes
+/// that could be read, instead of failing them all.
+///
+/// Errors in the structures that index the attributes (the Attribute Info
+/// message, the dense-storage heap header or B-tree) still fail the call:
+/// then it is unknown which attributes exist at all.
+pub fn extract_attributes_tolerant(
+    file_data: &[u8],
+    header: &ObjectHeader,
+    offset_size: u8,
+    length_size: u8,
+) -> Result<(Vec<AttributeMessage>, Vec<FormatError>), FormatError> {
+    extract_attributes_tolerant_core(file_data, header, offset_size, length_size)
+}
+
+/// [`extract_attributes_tolerant`] over any [`Storage`] (see
+/// [`extract_attributes_full_in`] for dense storage). One with the whole
+/// file in memory is read as the slice, by code compiled in this crate (see
+/// [`crate::storage`], "Slice entry points").
+#[inline]
+pub fn extract_attributes_tolerant_in<S: Storage + ?Sized>(
+    file_data: &S,
+    header: &ObjectHeader,
+    offset_size: u8,
+    length_size: u8,
+) -> Result<(Vec<AttributeMessage>, Vec<FormatError>), FormatError> {
+    match file_data.as_contiguous() {
+        Some(all) => extract_attributes_tolerant(all, header, offset_size, length_size),
+        None => extract_attributes_tolerant_core(file_data, header, offset_size, length_size),
+    }
+}
+
+fn extract_attributes_tolerant_core<S: Storage + ?Sized>(
+    file_data: &S,
+    header: &ObjectHeader,
+    offset_size: u8,
+    length_size: u8,
+) -> Result<(Vec<AttributeMessage>, Vec<FormatError>), FormatError> {
+    let mut errors = Vec::new();
+    let attrs = extract_attributes_with(file_data, header, offset_size, length_size, &mut |e| {
+        errors.push(e);
+        Ok(())
+    })?;
+    Ok((attrs, errors))
+}
+
+/// Read every attribute; each one that fails goes to `on_error`, which
+/// either stops the read (returns the error) or skips that attribute.
+fn extract_attributes_with<S: Storage + ?Sized>(
+    file_data: &S,
+    header: &ObjectHeader,
+    offset_size: u8,
+    length_size: u8,
+    on_error: &mut dyn FnMut(FormatError) -> Result<(), FormatError>,
+) -> Result<Vec<AttributeMessage>, FormatError> {
+    let mut attrs = Vec::new();
+    // Each attribute's creation order, where the file records one.
+    let mut orders: Vec<u32> = Vec::new();
+
+    extract_compact_attributes(
+        file_data,
+        header,
+        offset_size,
+        length_size,
+        &mut attrs,
+        &mut orders,
+        on_error,
+    )?;
+
+    // Check for dense attributes via AttributeInfo message
+    let attr_info = find_attribute_info(header, offset_size)?;
+    if let Some(info) = &attr_info
+        && let Some(fh_addr) = info.fractal_heap_address
+    {
+        extract_dense_attributes(
+            file_data,
+            info,
+            fh_addr,
+            offset_size,
+            length_size,
+            &mut attrs,
+            &mut orders,
+            on_error,
+        )?;
+    }
+
+    // An object that tracks attribute creation order lists its attributes
+    // in that order (h5py's `track_order=True`), as libhdf5 does; otherwise
+    // they come in storage order.
+    if attr_info.is_some_and(|i| i.max_creation_index.is_some()) {
+        let mut paired: Vec<(u32, AttributeMessage)> = orders.into_iter().zip(attrs).collect();
+        paired.sort_by_key(|(o, _)| *o);
+        attrs = paired.into_iter().map(|(_, a)| a).collect();
+    }
+
+    Ok(attrs)
+}
+
+/// B-tree v2 record type of dense attribute storage's name index.
+const ATTRIBUTE_NAME_INDEX: u8 = 8;
+
+/// The attribute called `name` on the object with header `header`: the
+/// first one [`extract_attributes_tolerant`] returns under that name, or
+/// `None` if it returns none (an attribute that cannot be read is not
+/// returned there either).
+///
+/// Compact attributes are in the header and are scanned. Dense attributes
+/// are found through the name index (a v2 B-tree of lookup3 name hashes,
+/// record type 8): only the attributes whose names hash like `name` are read
+/// from the heap, O(log n) instead of all of them. Errors in the structures
+/// that index the attributes fail the call, as they fail a listing.
+pub fn find_attribute_in_file(
+    file_data: &[u8],
+    header: &ObjectHeader,
+    name: &str,
+    offset_size: u8,
+    length_size: u8,
+) -> Result<Option<AttributeMessage>, FormatError> {
+    find_attribute_core(
+        file_data,
+        header,
+        name,
+        offset_size,
+        length_size,
+        &mut Vec::new(),
+    )
+}
+
+/// [`find_attribute_in_file`] over any [`Storage`] (see
+/// [`extract_attributes_full_in`] for dense storage, whose name index still
+/// needs the whole file in memory). One with the whole file in memory is
+/// read as the slice, by code compiled in this crate (see
+/// [`crate::storage`], "Slice entry points").
+#[inline]
+pub fn find_attribute_in<S: Storage + ?Sized>(
+    file_data: &S,
+    header: &ObjectHeader,
+    name: &str,
+    offset_size: u8,
+    length_size: u8,
+) -> Result<Option<AttributeMessage>, FormatError> {
+    match file_data.as_contiguous() {
+        Some(all) => find_attribute_in_file(all, header, name, offset_size, length_size),
+        None => find_attribute_core(
+            file_data,
+            header,
+            name,
+            offset_size,
+            length_size,
+            &mut Vec::new(),
+        ),
+    }
+}
+
+/// [`find_attribute_in`], also returning the errors of the attributes it
+/// could not read on the way (which it leaves out rather than failing
+/// the call): the attribute asked for may be one of them. A reader of a
+/// file that is being written uses them to tell a read that raced the
+/// writer from an absent attribute.
+pub fn find_attribute_reporting_in<S: Storage + ?Sized>(
+    file_data: &S,
+    header: &ObjectHeader,
+    name: &str,
+    offset_size: u8,
+    length_size: u8,
+) -> Result<(Option<AttributeMessage>, Vec<FormatError>), FormatError> {
+    let mut errors = Vec::new();
+    let found = find_attribute_core(
+        file_data,
+        header,
+        name,
+        offset_size,
+        length_size,
+        &mut errors,
+    )?;
+    Ok((found, errors))
+}
+
+fn find_attribute_core<S: Storage + ?Sized>(
+    file_data: &S,
+    header: &ObjectHeader,
+    name: &str,
+    offset_size: u8,
+    length_size: u8,
+    errors: &mut Vec<FormatError>,
+) -> Result<Option<AttributeMessage>, FormatError> {
+    let attr_info = find_attribute_info(header, offset_size)?;
+    let dense = attr_info
+        .as_ref()
+        .and_then(|i| Some((i.fractal_heap_address?, i.btree_name_index_address?)));
+    let Some((fh_addr, btree_addr)) = dense else {
+        // Compact only (or dense storage without a name index, which a
+        // listing reports): as a listing finds it.
+        let (attrs, errs) =
+            extract_attributes_tolerant_in(file_data, header, offset_size, length_size)?;
+        errors.extend(errs);
+        return Ok(attrs.into_iter().find(|a| a.name == name));
+    };
+    let btree_hdr = BTreeV2Header::parse_in(
+        file_data,
+        to_usize(btree_addr)? as u64,
+        offset_size,
+        length_size,
+    )?;
+    let fh = FractalHeapHeader::parse_in(file_data, fh_addr, offset_size, length_size)?;
+    if btree_hdr.tree_type != ATTRIBUTE_NAME_INDEX || btree_hdr.record_size < 4 {
+        let (attrs, errs) =
+            extract_attributes_tolerant_in(file_data, header, offset_size, length_size)?;
+        errors.extend(errs);
+        return Ok(attrs.into_iter().find(|a| a.name == name));
+    }
+
+    // A listing has the compact attributes first.
+    let mut compact = Vec::new();
+    extract_compact_attributes(
+        file_data,
+        header,
+        offset_size,
+        length_size,
+        &mut compact,
+        &mut Vec::new(),
+        &mut |_| Ok(()),
+    )?;
+    if let Some(a) = compact.into_iter().find(|a| a.name == name) {
+        return Ok(Some(a));
+    }
+
+    // Record: heap ID + message flags(1) + creation order(4) + hash(4); the
+    // hash is the last field.
+    let hash = jenkins_lookup3(name.as_bytes());
+    let hash_at = usize::from(btree_hdr.record_size) - 4;
+    let records = find_btree_v2_records_in(file_data, &btree_hdr, offset_size, &mut |r| match r
+        .get(hash_at..hash_at + 4)
+    {
+        Some(h) => u32::from_le_bytes([h[0], h[1], h[2], h[3]]).cmp(&hash),
+        None => core::cmp::Ordering::Less,
+    })?;
+    let id_len = usize::from(fh.heap_id_length);
+    for record in &records {
+        let Some(id_bytes) = record.data.get(..id_len) else {
+            continue;
+        };
+        let attr = fh
+            .read_managed_object_in(file_data, id_bytes, offset_size)
+            .and_then(|d| {
+                AttributeMessage::parse_in_storage(&d, file_data, offset_size, length_size)
+            });
+        // One that cannot be read is left out, as from a listing.
+        match attr {
+            Ok(attr) if attr.name == name => return Ok(Some(attr)),
+            Ok(_) => {}
+            Err(e) => errors.push(e),
+        }
+    }
+    Ok(None)
+}
+
+/// The attributes stored in the object header itself (compact storage), and
+/// each one's creation order into `orders`.
+fn extract_compact_attributes<S: Storage + ?Sized>(
+    file_data: &S,
+    header: &ObjectHeader,
+    offset_size: u8,
+    length_size: u8,
+    attrs: &mut Vec<AttributeMessage>,
+    orders: &mut Vec<u32>,
+    on_error: &mut dyn FnMut(FormatError) -> Result<(), FormatError>,
+) -> Result<(), FormatError> {
     for msg in &header.messages {
         if msg.msg_type == MessageType::Attribute {
-            if shared_message::is_shared(msg.flags) {
+            let attr = if shared_message::is_shared(msg.flags) {
                 // Shared attribute: resolve the reference to get actual attribute data
-                let shared_ref = shared_message::parse_shared_ref(&msg.data, offset_size)?;
-                let resolved_data = shared_message::resolve_shared_message(
+                shared_message::parse_shared_ref_sized(&msg.data, offset_size, length_size)
+                    .and_then(|shared_ref| {
+                        shared_message::resolve_shared_message_in(
                             file_data,
                             &shared_ref,
                             MessageType::Attribute,
                             offset_size,
                             length_size,
-                )?;
-                let attr = AttributeMessage::parse_in_file(
-                    &resolved_data,
+                        )
+                    })
+                    .and_then(|resolved| {
+                        AttributeMessage::parse_in_storage(
+                            &resolved,
                             file_data,
                             offset_size,
                             length_size,
-                )?;
-                attrs.push(attr);
+                        )
+                    })
             } else {
-                let attr = AttributeMessage::parse_in_file(
-                    &msg.data,
-                    file_data,
-                    offset_size,
-                    length_size,
-                )?;
+                AttributeMessage::parse_in_storage(&msg.data, file_data, offset_size, length_size)
+            };
+            let attr = attr.and_then(|a| check_in_header(a, header));
+            match attr {
+                Ok(attr) => {
                     attrs.push(attr);
+                    orders.push(msg.creation_order.map_or(0, u32::from));
+                }
+                Err(e) => on_error(e)?,
             }
         }
     }
-
-    // Check for dense attributes via AttributeInfo message
-    let attr_info = find_attribute_info(header, offset_size)?;
-    if let Some(info) = attr_info
-        && let Some(fh_addr) = info.fractal_heap_address
-    {
-        let dense_attrs =
-            extract_dense_attributes(file_data, &info, fh_addr, offset_size, length_size)?;
-        attrs.extend(dense_attrs);
-    }
-
-    Ok(attrs)
+    Ok(())
 }
 
 /// Find and parse the Attribute Info message from an object header.
@@ -461,16 +778,21 @@ fn find_attribute_info(
     Ok(None)
 }
 
-/// Extract attributes from dense storage (fractal heap + B-tree v2).
-fn extract_dense_attributes(
-    file_data: &[u8],
+/// Extract attributes from dense storage (fractal heap + B-tree v2), and
+/// each one's creation order into `orders`.
+#[allow(clippy::too_many_arguments)]
+fn extract_dense_attributes<S: Storage + ?Sized>(
+    file_data: &S,
     attr_info: &AttributeInfoMessage,
     fh_addr: u64,
     offset_size: u8,
     length_size: u8,
-) -> Result<Vec<AttributeMessage>, FormatError> {
+    attrs: &mut Vec<AttributeMessage>,
+    orders: &mut Vec<u32>,
+    on_error: &mut dyn FnMut(FormatError) -> Result<(), FormatError>,
+) -> Result<(), FormatError> {
     // Parse fractal heap
-    let fh = FractalHeapHeader::parse(file_data, fh_addr as usize, offset_size, length_size)?;
+    let fh = FractalHeapHeader::parse_in(file_data, fh_addr, offset_size, length_size)?;
 
     // Parse B-tree v2 for name index (type 8)
     let btree_addr = attr_info
@@ -479,31 +801,47 @@ fn extract_dense_attributes(
             expected: 1,
             available: 0,
         })?;
-    let btree_hdr = BTreeV2Header::parse(file_data, btree_addr as usize, offset_size, length_size)?;
-    let records = collect_btree_v2_records(file_data, &btree_hdr, offset_size, length_size)?;
+    let btree_hdr = BTreeV2Header::parse_in(
+        file_data,
+        to_usize(btree_addr)? as u64,
+        offset_size,
+        length_size,
+    )?;
+    let records = collect_btree_v2_records_in(file_data, &btree_hdr, offset_size, length_size)?;
 
-    let mut attrs = Vec::new();
     for record in &records {
         // Per HDF5 spec, both type 8 and type 9 records start with heap_id:
         //   Type 8: heap_id(8) + msg_flags(1) + creation_order(4) + hash(4)
         //   Type 9: heap_id(8) + msg_flags(1) + creation_order(4)
-        let id_offset = 0;
-
-        if record.data.len() < id_offset + fh.heap_id_length as usize {
+        let id_len = fh.heap_id_length as usize;
+        let Some(id_bytes) = record.data.get(..id_len) else {
+            on_error(FormatError::UnexpectedEof {
+                expected: id_len,
+                available: record.data.len(),
+            })?;
             continue;
-        }
-        let id_bytes = &record.data[id_offset..id_offset + fh.heap_id_length as usize];
-
-        // Read attribute message from fractal heap
-        let attr_data = fh.read_managed_object(file_data, id_bytes, offset_size)?;
+        };
 
         // The data in the heap is a complete attribute message
-        let attr =
-            AttributeMessage::parse_in_file(&attr_data, file_data, offset_size, length_size)?;
+        let attr = fh
+            .read_managed_object_in(file_data, id_bytes, offset_size)
+            .and_then(|attr_data| {
+                AttributeMessage::parse_in_storage(&attr_data, file_data, offset_size, length_size)
+            });
+        match attr {
+            Ok(attr) => {
                 attrs.push(attr);
+                let order = record
+                    .data
+                    .get(id_len + 1..id_len + 5)
+                    .map_or(0, |b| u32::from_le_bytes([b[0], b[1], b[2], b[3]]));
+                orders.push(order);
+            }
+            Err(e) => on_error(e)?,
+        }
     }
 
-    Ok(attrs)
+    Ok(())
 }
 
 #[cfg(test)]
@@ -523,7 +861,8 @@ mod tests {
 
     /// Build an f64 LE datatype message.
     fn build_f64_dt() -> Vec<u8> {
-        let mut buf = build_dt_header(1, 1, [0x00, 0x00, 0x02], 8);
+        // Sign bit 63 (bits 8-15 of the class bits).
+        let mut buf = build_dt_header(1, 1, [0x20, 63, 0x00], 8);
         let mut props = [0u8; 12];
         props[2..4].copy_from_slice(&64u16.to_le_bytes()); // bit_precision
         props[4] = 52; // exp_location
@@ -897,4 +1236,73 @@ mod tests {
         let strs = attr.read_as_strings().unwrap();
         assert_eq!(strs, vec!["abcd", "EFGH"]);
     }
+
+    /// Every object's attributes in h5py-written files read identically
+    /// through a read_at-only CountingStorage — compact ones, shared ones,
+    /// those behind an Attribute Info message and dense storage (its v2
+    /// B-tree name index included) — and through a slice as Storage.
+    #[test]
+    fn storage_reads_match_slice_reads() {
+        use crate::storage::CountingStorage;
+        let files: [(&str, &[u8]); 5] = [
+            ("attrs", include_bytes!("../tests/fixtures/attrs.h5")),
+            (
+                "mixed_attrs",
+                include_bytes!("../tests/fixtures/mixed_attrs.h5"),
+            ),
+            (
+                "dense_attrs",
+                include_bytes!("../tests/fixtures/dense_attrs.h5"),
+            ),
+            (
+                "dense_attrs_root",
+                include_bytes!("../tests/fixtures/dense_attrs_root.h5"),
+            ),
+            (
+                "shared_fill_value",
+                include_bytes!("../tests/fixtures/shared_fill_value.h5"),
+            ),
+        ];
+        let (mut same, mut dense, mut attrs) = (0, 0, 0);
+        for (name, file) in files {
+            let sb = crate::superblock::Superblock::parse(file, 0).unwrap();
+            let (os, ls) = (sb.offset_size, sb.length_size);
+            let mut addrs = vec![sb.root_group_address];
+            addrs.extend(
+                crate::group_v2::resolve_group_children(file, &sb, sb.root_group_address)
+                    .unwrap()
+                    .iter()
+                    .map(|e| e.object_header_address),
+            );
+            let storage = CountingStorage::new(file.to_vec());
+            for addr in addrs {
+                let header = ObjectHeader::parse(file, addr as usize, os, ls).unwrap();
+                let want = extract_attributes_full(file, &header, os, ls);
+                let slice_storage = extract_attributes_full_in(&file, &header, os, ls);
+                assert_eq!(format!("{slice_storage:?}"), format!("{want:?}"));
+                let got = extract_attributes_full_in(&storage, &header, os, ls);
+                let got_t = extract_attributes_tolerant_in(&storage, &header, os, ls);
+                let is_dense = find_attribute_info(&header, os)
+                    .unwrap()
+                    .is_some_and(|i| i.fractal_heap_address.is_some());
+                if is_dense {
+                    dense += 1;
+                }
+                attrs += want.as_ref().map_or(0, Vec::len);
+                assert_eq!(format!("{got:?}"), format!("{want:?}"), "{name}");
+                let want_t = extract_attributes_tolerant(file, &header, os, ls);
+                assert_eq!(format!("{got_t:?}"), format!("{want_t:?}"), "{name}");
+                same += 1;
+                for a in want.iter().flatten() {
+                    let one = find_attribute_in(&storage, &header, &a.name, os, ls);
+                    let want_one = find_attribute_in_file(file, &header, &a.name, os, ls);
+                    assert_eq!(format!("{one:?}"), format!("{want_one:?}"), "{name}");
+                }
+            }
+        }
+        assert!(
+            same >= 5 && dense >= 2 && attrs >= 5,
+            "{same} {dense} {attrs}"
+        );
+    }
 }

crates/clawhdf5-format/src/btree_v1.rs

@@ -4,6 +4,7 @@
 use alloc::vec::Vec;
 
 use crate::error::FormatError;
+use crate::storage::{Storage, read_exact_at};
 
 /// A parsed B-tree v1 node.
 #[derive(Debug, Clone)]
@@ -74,13 +75,30 @@ impl BTreeV1Node {
         file_data: &[u8],
         offset: usize,
         offset_size: u8,
+        length_size: u8,
+    ) -> Result<BTreeV1Node, FormatError> {
+        Self::parse_in(file_data, offset as u64, offset_size, length_size)
+    }
+
+    /// [`Self::parse`] over any [`Storage`]: one read of the node's header,
+    /// one of its keys and children.
+    pub fn parse_in<S: Storage + ?Sized>(
+        file: &S,
+        offset: u64,
+        offset_size: u8,
         _length_size: u8,
     ) -> Result<BTreeV1Node, FormatError> {
         // signature(4) + node_type(1) + node_level(1) + entries_used(2) = 8
         // + left_sibling(offset_size) + right_sibling(offset_size)
         let os = offset_size as usize;
         let header_size = 8 + os * 2;
-        ensure_len(file_data, offset, header_size)?;
+        // The body is read once the header says how long it is.
+        file.hint(offset, NODE_HINT_LEN);
+        let header = read_exact_at(file, offset, header_size)?;
+        let file_data: &[u8] = &header;
+        // The header's read checked that `offset + header_size` fits.
+        let body_start = offset + header_size as u64;
+        let offset = 0usize;
 
         if &file_data[offset..offset + 4] != b"TREE" {
             return Err(FormatError::InvalidBTreeSignature);
@@ -102,31 +120,30 @@ impl BTreeV1Node {
         } else {
             Some(read_offset(file_data, pos, offset_size)?)
         };
-        pos += os;
 
         // For type 0: keys are offset_size bytes, children are offset_size bytes
         // Layout: key[0], child[0], key[1], child[1], ..., key[N-1], child[N-1], key[N]
         let eu = entries_used as usize;
         let key_size = os; // For type 0, key = offset_size
         let needed = eu * (key_size + os) + key_size; // eu children + (eu+1) keys
-        ensure_len(file_data, pos, needed)?;
+        let body = read_exact_at(file, body_start, needed)?;
+        let file_data: &[u8] = &body;
 
         let mut keys = Vec::with_capacity(eu + 1);
         let mut children = Vec::with_capacity(eu);
 
-        for _i in 0..eu {
-            // key[i]
-            let key = read_offset(file_data, pos, offset_size)?;
-            keys.push(key);
-            pos += key_size;
-            // child[i]
-            let child = read_offset(file_data, pos, offset_size)?;
-            children.push(child);
-            pos += os;
+        if os == 0 {
+            // What reading the first key reports (and keeps `chunks_exact`
+            // below from being given a zero size).
+            return Err(FormatError::InvalidOffsetSize(offset_size));
         }
-        // final key
-        let key = read_offset(file_data, pos, offset_size)?;
-        keys.push(key);
+        // `needed` bytes: key[0], child[0], ..., child[eu - 1], key[eu].
+        let (pairs, last) = file_data.split_at(eu * (key_size + os));
+        for pair in pairs.chunks_exact(key_size + os) {
+            keys.push(read_offset(pair, 0, offset_size)?);
+            children.push(read_offset(pair, key_size, offset_size)?);
+        }
+        keys.push(read_offset(last, 0, offset_size)?);
 
         Ok(BTreeV1Node {
             node_type,
@@ -141,7 +158,18 @@ impl BTreeV1Node {
 }
 
 /// Maximum recursion depth for B-tree traversal (malformed data protection).
-const MAX_BTREE_DEPTH: usize = 64;
+pub(crate) const MAX_BTREE_DEPTH: usize = 64;
+
+/// What a symbol table node takes with libhdf5's default group leaf K (4):
+/// its 8-byte header and 2K entries of 40 bytes (8-byte offsets). Hinted
+/// before one is read ([`Storage::hint`]); a node of another size is read
+/// all the same.
+const SNOD_HINT_LEN: usize = 8 + 8 * 40;
+
+/// What a group B-tree node takes with libhdf5's default internal K (16):
+/// its header (24 bytes with 8-byte offsets), 2K + 1 keys and 2K children
+/// of 8 bytes. Hinted before one is read.
+const NODE_HINT_LEN: usize = 24 + (2 * 16 + 1 + 2 * 16) * 8;
 
 /// Collect all leaf-level child addresses (SNOD addresses) by traversing the B-tree.
 pub fn collect_symbol_table_nodes(
@@ -150,11 +178,21 @@ pub fn collect_symbol_table_nodes(
     offset_size: u8,
     length_size: u8,
 ) -> Result<Vec<u64>, FormatError> {
-    collect_symbol_table_nodes_inner(file_data, btree_address, offset_size, length_size, 0)
+    collect_symbol_table_nodes_in(file_data, btree_address, offset_size, length_size)
 }
 
-fn collect_symbol_table_nodes_inner(
-    file_data: &[u8],
+/// [`collect_symbol_table_nodes`] over any [`Storage`]: two reads per node.
+pub fn collect_symbol_table_nodes_in<S: Storage + ?Sized>(
+    file: &S,
+    btree_address: u64,
+    offset_size: u8,
+    length_size: u8,
+) -> Result<Vec<u64>, FormatError> {
+    collect_symbol_table_nodes_inner(file, btree_address, offset_size, length_size, 0)
+}
+
+fn collect_symbol_table_nodes_inner<S: Storage + ?Sized>(
+    file: &S,
     btree_address: u64,
     offset_size: u8,
     length_size: u8,
@@ -164,29 +202,47 @@ fn collect_symbol_table_nodes_inner(
         return Err(FormatError::NestingDepthExceeded);
     }
 
-    let node = BTreeV1Node::parse(file_data, btree_address as usize, offset_size, length_size)?;
+    let node = BTreeV1Node::parse_in(file, btree_address, offset_size, length_size)?;
 
     if node.node_type != 0 {
         return Err(FormatError::InvalidBTreeNodeType(node.node_type));
     }
 
     if node.node_level == 0 {
-        // Leaf: children are SNOD addresses
+        // Leaf: children are SNOD addresses, read next (see
+        // `Storage::hint`).
+        for &snod in &node.children {
+            file.hint(snod, SNOD_HINT_LEN);
+        }
         Ok(node.children)
     } else {
-        // Internal: recurse into children
+        // Internal: recurse into children. A child that fails does not
+        // stop the walk: the others are still descended into (reading, not
+        // using, what they hold), then the first error is returned. The
+        // result and the error are those of stopping at the first failure;
+        // a storage that records what it lacks (see `storage::touch`)
+        // learns every node the walk can reach in one attempt.
         let mut result = Vec::new();
+        let mut failed = None;
         for &child_addr in &node.children {
-            let child_snods = collect_symbol_table_nodes_inner(
-                file_data,
+            match collect_symbol_table_nodes_inner(
+                file,
                 child_addr,
                 offset_size,
                 length_size,
                 depth + 1,
-            )?;
-            result.extend(child_snods);
+            ) {
+                Ok(child_snods) if failed.is_none() => result.extend(child_snods),
+                Ok(_) => {}
+                Err(e) => {
+                    failed.get_or_insert(e);
+                }
+            }
+        }
+        match failed {
+            Some(e) => Err(e),
+            None => Ok(result),
         }
-        Ok(result)
     }
 }
 
@@ -317,4 +373,48 @@ mod tests {
         assert_eq!(node.entries_used, 1);
         assert_eq!(node.children, vec![0x50]);
     }
+
+    /// Nodes and trees, cut at every length, parse identically through a
+    /// `read_at`-only storage.
+    #[test]
+    fn storage_parse_matches_slice_parse() {
+        use crate::storage::CountingStorage;
+        let nodes = [
+            build_btree_node(0, 0, &[0, 5, 10], &[0x100, 0x200], None, None, 8),
+            build_btree_node(0, 0, &[0, 5], &[0x100], Some(0x40), Some(0x80), 4),
+            build_btree_node(1, 2, &[0, 5], &[0x100], None, Some(0x80), 8),
+        ];
+        for (n, node) in nodes.iter().enumerate() {
+            let os = if n == 1 { 4 } else { 8 };
+            for cut in 0..=node.len() {
+                let f = &node[..cut];
+                let storage = CountingStorage::new(f.to_vec());
+                let want = BTreeV1Node::parse(f, 0, os, 8);
+                let got = BTreeV1Node::parse_in(&storage, 0, os, 8);
+                assert_eq!(format!("{got:?}"), format!("{want:?}"));
+            }
+        }
+        let leaf1 = build_btree_node(0, 0, &[0, 5], &[0xA00], None, None, 8);
+        let leaf2 = build_btree_node(0, 0, &[5, 10], &[0xB00], None, None, 8);
+        let internal = build_btree_node(0, 1, &[0, 5, 10], &[0, 256], None, None, 8);
+        let mut file = vec![0u8; 512 + internal.len()];
+        file[..leaf1.len()].copy_from_slice(&leaf1);
+        file[256..256 + leaf2.len()].copy_from_slice(&leaf2);
+        file[512..].copy_from_slice(&internal);
+        for cut in [file.len(), 300, 260, 100, 10] {
+            let mut f = file.clone();
+            if cut < 512 {
+                // Truncate the leaves, keep the root.
+                f[cut..512].fill(0);
+            }
+            let storage = CountingStorage::new(f.clone());
+            assert_eq!(
+                collect_symbol_table_nodes_in(&storage, 512, 8, 8),
+                collect_symbol_table_nodes(&f, 512, 8, 8)
+            );
+        }
+        let storage = CountingStorage::new(file);
+        collect_symbol_table_nodes_in(&storage, 512, 8, 8).unwrap();
+        assert_eq!(storage.reads(), 6);
+    }
 }

crates/clawhdf5-format/src/btree_v1_write.rs

@@ -0,0 +1,470 @@
+//! Writing a version-1 B-tree chunk index (node type 1): the chunk index of
+//! layout message versions 1-3, and the only one HDF5 1.8 reads.
+//!
+//! The tree is built the way libhdf5 builds it when the chunks reach it one
+//! after another in row-major order (a whole-dataset `H5Dwrite` of a 1-D
+//! dataset, or of any dataset without a chunk cache; with one, libhdf5
+//! inserts the small chunks of a multi-dimensional dataset in the order its
+//! cache evicts them, which fills the nodes differently): each
+//! chunk goes through the same steps as `H5B_insert` (`H5B.c`) with the
+//! chunk callbacks of `H5Dbtree.c`, so nodes split where libhdf5's split,
+//! with its default split ratios (a full right-most node keeps 90% of its
+//! children, a left-most one 10%, any other half), and keys hold what
+//! libhdf5's hold:
+//!
+//! - a chunk's key is its size in the file, its filter mask and its offsets
+//!   (the element-size coordinate 0);
+//! - a node's final key is the zero-size key one chunk past the chunk that
+//!   last moved it (every scaled coordinate plus one, `H5D__btree_new_node`),
+//!   which libhdf5 moves only when a new chunk is not below it
+//!   (`H5D__btree_cmp3`) — so after an even number of appends in one
+//!   dimension it lies on the last chunk itself;
+//! - a full root is copied to a new node and becomes the parent of the copy
+//!   and its new sibling, so the root's address (the layout message's) never
+//!   changes.
+//!
+//! Nodes are laid out in the order libhdf5 allocates them (the root first,
+//! then each new node as a split creates it), all of the full node size, the
+//! unused slots zero.
+
+#[cfg(not(feature = "std"))]
+use alloc::{format, vec, vec::Vec};
+
+use core::cmp::Ordering;
+
+use crate::error::FormatError;
+
+/// libhdf5's default chunk B-tree K (`HDF5_BTREE_CHUNK_IK_DEF`): nodes hold
+/// up to 2K = 64 children. Superblocks of version 2 cannot record another
+/// value without a superblock extension, which this writer does not emit.
+pub(crate) const CHUNK_BTREE_K: u16 = 32;
+
+/// libhdf5's default split ratios (`H5D_XFER_BTREE_SPLIT_RATIO_DEF`) for a
+/// left-most, middle and right-most node.
+const SPLIT_RATIOS: [f64; 3] = [0.1, 0.5, 0.9];
+
+/// A chunk to index: scaled coordinates (offset / chunk dimension) in each
+/// dataset dimension, stored size, filter mask and address.
+pub(crate) struct ChunkEntry {
+    pub(crate) scaled: Vec<u64>,
+    pub(crate) nbytes: u64,
+    pub(crate) filter_mask: u32,
+    pub(crate) address: u64,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq)]
+struct Key {
+    nbytes: u32,
+    mask: u32,
+    /// Scaled coordinates, the element-size one (0 or 1) last.
+    scaled: Vec<u64>,
+}
+
+impl Key {
+    /// `H5D__btree_new_node`'s right key: one chunk past `self` in every
+    /// dimension, with no storage.
+    fn right_of(&self) -> Key {
+        Key {
+            nbytes: 0,
+            mask: 0,
+            scaled: self.scaled.iter().map(|s| s + 1).collect(),
+        }
+    }
+
+    fn cmp_scaled(&self, other: &Key) -> Ordering {
+        self.scaled.cmp(&other.scaled)
+    }
+}
+
+#[derive(Debug, Clone)]
+struct Node {
+    level: u8,
+    left: Option<usize>,
+    right: Option<usize>,
+    /// `children.len() + 1` keys once the node holds a child.
+    keys: Vec<Key>,
+    /// Chunk addresses in a leaf, node indexes above.
+    children: Vec<u64>,
+}
+
+/// What an insertion below a node did (`H5B__insert_helper`'s outputs).
+#[derive(Default)]
+struct Ret {
+    /// The node's new left key (`lt_key_changed`).
+    lt: Option<Key>,
+    /// The node's new right key (`rt_key_changed`).
+    rt: Option<Key>,
+    /// The node split: the key shared by the halves and the new right node.
+    split: Option<(Key, usize)>,
+}
+
+struct Tree {
+    nodes: Vec<Node>,
+    two_k: usize,
+}
+
+fn bad(why: &str) -> FormatError {
+    FormatError::SerializationError(format!("version-1 B-tree chunk index: {why}"))
+}
+
+impl Tree {
+    fn new(k: u16) -> Self {
+        Self {
+            nodes: vec![Node {
+                level: 0,
+                left: None,
+                right: None,
+                keys: Vec::new(),
+                children: Vec::new(),
+            }],
+            two_k: 2 * usize::from(k),
+        }
+    }
+
+    /// `H5B_insert` of `key` (a chunk after every chunk already inserted).
+    fn insert(&mut self, key: &Key, addr: u64) -> Result<(), FormatError> {
+        let r = self.insert_helper(0, key, addr, 64)?;
+        let Some((md, split)) = r.split else {
+            return Ok(());
+        };
+        // The root split: copy it to a new node and make the root the
+        // parent of the copy and its new right sibling.
+        let lt = r.lt.unwrap_or_else(|| self.nodes[0].keys[0].clone());
+        let rt = match r.rt {
+            Some(rt) => rt,
+            None => self.nodes[split]
+                .keys
+                .last()
+                .cloned()
+                .ok_or_else(|| bad("empty node"))?,
+        };
+        let moved = self.nodes[0].clone();
+        let level = moved.level;
+        let moved_id = self.nodes.len();
+        self.nodes.push(moved);
+        self.nodes[split].left = Some(moved_id);
+        self.nodes[0] = Node {
+            level: level + 1,
+            left: None,
+            right: None,
+            keys: vec![lt, md, rt],
+            children: vec![moved_id as u64, split as u64],
+        };
+        Ok(())
+    }
+
+    fn insert_helper(
+        &mut self,
+        id: usize,
+        key: &Key,
+        addr: u64,
+        depth: u8,
+    ) -> Result<Ret, FormatError> {
+        if depth == 0 {
+            return Err(bad("tree too deep"));
+        }
+        let n = self.nodes[id].children.len();
+        let level = self.nodes[id].level;
+        let mut ret = Ret::default();
+        if n == 0 {
+            // The first chunk (H5B_INS_FIRST): its key and the right key.
+            let node = &mut self.nodes[id];
+            node.keys = vec![key.clone(), key.right_of()];
+            node.children = vec![addr];
+            return Ok(ret);
+        }
+        // Binary search with H5D__btree_cmp3: 1 when the chunk is not below
+        // the right key, -1 when below the left key, else 0.
+        let (mut lo, mut hi, mut idx) = (0usize, n, 0usize);
+        let mut cmp = Ordering::Less;
+        while lo < hi && cmp != Ordering::Equal {
+            idx = (lo + hi) / 2;
+            let node = &self.nodes[id];
+            cmp = if key.cmp_scaled(&node.keys[idx + 1]) != Ordering::Less {
+                Ordering::Greater
+            } else if key.cmp_scaled(&node.keys[idx]) == Ordering::Less {
+                Ordering::Less
+            } else {
+                Ordering::Equal
+            };
+            if cmp == Ordering::Less {
+                hi = idx;
+            } else {
+                lo = idx + 1;
+            }
+        }
+        let (mut lt_changed, mut rt_changed) = (false, false);
+        // The child to add after child `idx`, with its left key.
+        let mut new_child: Option<(Key, u64)> = None;
+        match cmp {
+            Ordering::Less => return Err(bad("chunks out of order")),
+            Ordering::Greater if idx + 1 < n => {
+                return Err(bad("cannot place chunk"));
+            }
+            Ordering::Greater if level == 0 => {
+                // Past every chunk of the right-most leaf: a new maximum
+                // (H5B_INS_RIGHT through `new_node`), which moves the right
+                // key one chunk past it.
+                idx = n - 1;
+                self.nodes[id].keys[idx + 1] = key.right_of();
+                rt_changed = true;
+                new_child = Some((key.clone(), addr));
+            }
+            Ordering::Equal if level == 0 => {
+                // Inside the last chunk's range: H5D__btree_insert adds it
+                // to the right of that chunk; the right key stays.
+                if key.scaled == self.nodes[id].keys[idx].scaled {
+                    return Err(bad("duplicate chunk"));
+                }
+                new_child = Some((key.clone(), addr));
+            }
+            _ => {
+                if cmp == Ordering::Greater {
+                    idx = n - 1;
+                }
+                let child = usize::try_from(self.nodes[id].children[idx])
+                    .map_err(|_| bad("bad node index"))?;
+                let r = self.insert_helper(child, key, addr, depth - 1)?;
+                if let Some(lt) = r.lt {
+                    self.nodes[id].keys[idx] = lt;
+                    lt_changed = true;
+                }
+                if let Some(rt) = r.rt {
+                    self.nodes[id].keys[idx + 1] = rt;
+                    rt_changed = true;
+                }
+                if let Some((md, split)) = r.split {
+                    new_child = Some((md, split as u64));
+                }
+            }
+        }
+        // Pass the node's changed end keys up, as H5B__insert_helper does.
+        if lt_changed && idx == 0 {
+            ret.lt = Some(self.nodes[id].keys[0].clone());
+        }
+        if rt_changed && idx + 1 >= n {
+            ret.rt = Some(self.nodes[id].keys[idx + 1].clone());
+        }
+        if let Some((md, child)) = new_child {
+            // A full node splits first; the child goes to the half that
+            // holds child `idx`.
+            let (mut target, mut split) = (id, None);
+            if n == self.two_k {
+                let s = self.split(id, idx);
+                let nleft = self.nodes[id].children.len();
+                if idx >= nleft {
+                    idx -= nleft;
+                    target = s;
+                }
+                split = Some(s);
+            }
+            // H5B__insert_child (H5B_INS_RIGHT): the new child after child
+            // `idx`, its left key after that child's.
+            let node = &mut self.nodes[target];
+            node.keys.insert(idx + 1, md);
+            node.children.insert(idx + 1, child);
+            ret.split = split.map(|s| (self.nodes[s].keys[0].clone(), s));
+        }
+        Ok(ret)
+    }
+
+    /// `H5B__split` of the full node `id`, the insertion going after child
+    /// `idx`; returns the new right node.
+    fn split(&mut self, id: usize, idx: usize) -> usize {
+        let node = &self.nodes[id];
+        let ratio = if node.right.is_none() {
+            SPLIT_RATIOS[2]
+        } else if node.left.is_none() {
+            SPLIT_RATIOS[0]
+        } else {
+            SPLIT_RATIOS[1]
+        };
+        let mut nleft = (self.two_k as f64 * ratio) as usize;
+        if idx < nleft && nleft == self.two_k {
+            nleft -= 1;
+        } else if idx >= nleft && nleft == 0 {
+            nleft += 1;
+        }
+        let new_id = self.nodes.len();
+        let right = Node {
+            level: node.level,
+            left: Some(id),
+            right: node.right,
+            keys: node.keys[nleft..].to_vec(),
+            children: node.children[nleft..].to_vec(),
+        };
+        let old_right = node.right;
+        self.nodes.push(right);
+        if let Some(r) = old_right {
+            self.nodes[r].left = Some(new_id);
+        }
+        let node = &mut self.nodes[id];
+        node.keys.truncate(nleft + 1);
+        node.children.truncate(nleft);
+        node.right = Some(new_id);
+        new_id
+    }
+}
+
+/// Bytes of one node of a chunk B-tree with `ndims` key dimensions (the
+/// dataset's rank plus the element-size one).
+fn node_size(two_k: usize, ndims: usize, offset_size: usize) -> usize {
+    let key = 8 + 8 * ndims;
+    8 + 2 * offset_size + (two_k + 1) * key + two_k * offset_size
+}
+
+/// Build the chunk B-tree for `chunks`, given in row-major order of their
+/// scaled coordinates, with nodes laid out from `base_address`. `chunk_dims`
+/// are the chunk's dimensions (the dataset's rank of them) and `elem_size`
+/// the element size, the key's last dimension. Returns the nodes' bytes; the
+/// root is at `base_address`. `chunks` must not be empty: an index without
+/// chunks has no tree (its address is undefined).
+pub(crate) fn build_chunk_btree_v1_at(
+    chunks: &[ChunkEntry],
+    chunk_dims: &[u64],
+    elem_size: u32,
+    base_address: u64,
+    offset_size: u8,
+) -> Result<Vec<u8>, FormatError> {
+    if chunks.is_empty() {
+        return Err(bad("no chunks"));
+    }
+    let rank = chunk_dims.len();
+    let mut tree = Tree::new(CHUNK_BTREE_K);
+    for c in chunks {
+        if c.scaled.len() != rank {
+            return Err(bad("chunk rank differs from the dataset's"));
+        }
+        let nbytes = u32::try_from(c.nbytes).map_err(|_| {
+            FormatError::SerializationError(format!(
+                "a chunk of {} bytes cannot be indexed by a version-1 B-tree \
+                 (HDF5 1.8 chunks are under 4 GiB)",
+                c.nbytes
+            ))
+        })?;
+        let mut scaled = c.scaled.clone();
+        scaled.push(0);
+        let key = Key {
+            nbytes,
+            mask: c.filter_mask,
+            scaled,
+        };
+        tree.insert(&key, c.address)?;
+    }
+
+    let os = usize::from(offset_size);
+    let ndims = rank + 1;
+    let nsize = node_size(tree.two_k, ndims, os);
+    let addr_of = |id: usize| base_address + (id * nsize) as u64;
+    let mut dims: Vec<u64> = chunk_dims.to_vec();
+    dims.push(u64::from(elem_size));
+    let mut out = vec![0u8; tree.nodes.len() * nsize];
+    for (i, node) in tree.nodes.iter().enumerate() {
+        let d = &mut out[i * nsize..(i + 1) * nsize];
+        d[0..4].copy_from_slice(b"TREE");
+        d[4] = 1; // node type: raw data chunks
+        d[5] = node.level;
+        let n = u16::try_from(node.children.len()).map_err(|_| bad("node too large"))?;
+        d[6..8].copy_from_slice(&n.to_le_bytes());
+        let undef = u64::MAX;
+        put_addr(&mut d[8..], node.left.map_or(undef, addr_of), os);
+        put_addr(&mut d[8 + os..], node.right.map_or(undef, addr_of), os);
+        let mut p = 8 + 2 * os;
+        for (k, key) in node.keys.iter().enumerate() {
+            d[p..p + 4].copy_from_slice(&key.nbytes.to_le_bytes());
+            d[p + 4..p + 8].copy_from_slice(&key.mask.to_le_bytes());
+            for (j, (&s, &dim)) in key.scaled.iter().zip(&dims).enumerate() {
+                let off = s
+                    .checked_mul(dim)
+                    .ok_or_else(|| FormatError::Overflow("chunk key offset".into()))?;
+                d[p + 8 + 8 * j..p + 16 + 8 * j].copy_from_slice(&off.to_le_bytes());
+            }
+            p += 8 + 8 * ndims;
+            if let Some(&child) = node.children.get(k) {
+                let a = if node.level == 0 {
+                    child
+                } else {
+                    addr_of(usize::try_from(child).map_err(|_| bad("bad node index"))?)
+                };
+                put_addr(&mut d[p..], a, os);
+                p += os;
+            }
+        }
+    }
+    Ok(out)
+}
+
+fn put_addr(d: &mut [u8], v: u64, os: usize) {
+    d[..os].copy_from_slice(&v.to_le_bytes()[..os]);
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn build(n: u64) -> Tree {
+        let mut t = Tree::new(CHUNK_BTREE_K);
+        for i in 0..n {
+            let key = Key {
+                nbytes: 80,
+                mask: 0,
+                scaled: vec![i, 0],
+            };
+            t.insert(&key, 1000 + i).unwrap();
+        }
+        t
+    }
+
+    /// Leaves in order from the root, with their child counts.
+    fn leaves(t: &Tree, id: usize, out: &mut Vec<usize>) {
+        let n = &t.nodes[id];
+        if n.level == 0 {
+            out.push(n.children.len());
+        } else {
+            for &c in &n.children {
+                leaves(t, c as usize, out);
+            }
+        }
+    }
+
+    #[test]
+    fn sequential_appends_split_as_libhdf5_does() {
+        // libhdf5 2.0 (h5py, libver=('v108', 'latest')) writes 1000 chunks
+        // as a root over 17 leaves of 57 chunks and one of 31, with the
+        // root's right key on the last chunk (9990, 8 for 10-element f8
+        // chunks).
+        let t = build(1000);
+        assert_eq!(t.nodes[0].level, 1);
+        let mut l = Vec::new();
+        leaves(&t, 0, &mut l);
+        let mut want = vec![57; 17];
+        want.push(31);
+        assert_eq!(l, want);
+        assert_eq!(t.nodes[0].keys.last().unwrap().scaled, vec![999, 1]);
+        // 100 000 chunks: three levels, a root of 31 children.
+        let t = build(100_000);
+        assert_eq!(t.nodes[0].level, 2);
+        assert_eq!(t.nodes[0].children.len(), 31);
+    }
+
+    #[test]
+    fn right_key_moves_every_other_append() {
+        let t = build(5);
+        assert_eq!(t.nodes[0].keys.last().unwrap().scaled, vec![5, 1]);
+        let t = build(6);
+        assert_eq!(t.nodes[0].keys.last().unwrap().scaled, vec![5, 1]);
+    }
+
+    #[test]
+    fn keys_and_siblings_are_consistent() {
+        let t = build(5000);
+        for (i, n) in t.nodes.iter().enumerate() {
+            assert!(n.children.len() <= t.two_k);
+            assert_eq!(n.keys.len(), n.children.len() + 1);
+            if let Some(r) = n.right {
+                assert_eq!(t.nodes[r].left, Some(i));
+                assert_eq!(n.keys.last(), t.nodes[r].keys.first());
+            }
+        }
+    }
+}

crates/clawhdf5-format/src/btree_v2.rs

@@ -2,11 +2,14 @@
 
 #[cfg(not(feature = "std"))]
 use alloc::vec::Vec;
+use core::cmp::Ordering;
 
 #[cfg(feature = "checksum")]
 use byteorder::{ByteOrder, LittleEndian};
 
+use crate::addr::to_usize;
 use crate::error::FormatError;
+use crate::storage::{Storage, Window, len_usize};
 
 /// Parsed B-tree v2 header (signature "BTHD").
 #[derive(Debug, Clone)]
@@ -71,7 +74,7 @@ fn ensure_len(data: &[u8], pos: usize, needed: usize) -> Result<(), FormatError>
 
 /// Compute the number of bytes needed to represent a count, using variable-width encoding.
 /// B-tree v2 uses this for the number of records fields in internal nodes.
-fn bytes_for_max_records(max_nrec: u64) -> usize {
+pub(crate) fn bytes_for_max_records(max_nrec: u64) -> usize {
     if max_nrec == 0 {
         return 1;
     }
@@ -97,38 +100,52 @@ impl BTreeV2Header {
         offset_size: u8,
         length_size: u8,
     ) -> Result<BTreeV2Header, FormatError> {
-        ensure_len(file_data, offset, 4)?;
-        if &file_data[offset..offset + 4] != b"BTHD" {
+        Self::parse_in(file_data, offset as u64, offset_size, length_size)
+    }
+
+    /// [`Self::parse`] over any [`Storage`]: one bounded read of the
+    /// header.
+    pub fn parse_in<S: Storage + ?Sized>(
+        file: &S,
+        offset: u64,
+        offset_size: u8,
+        length_size: u8,
+    ) -> Result<BTreeV2Header, FormatError> {
+        // Every field and the checksum; the window holds all of it or ends
+        // at the end of the file, so its bounds checks are the whole-file
+        // ones.
+        let full = 16 + usize::from(offset_size) + 2 + usize::from(length_size) + 4;
+        let w = Window::read(file, offset, full)?;
+        let d = &w.bytes;
+        w.ensure(0, 4)?;
+        if &d[..4] != b"BTHD" {
             return Err(FormatError::InvalidBTreeV2Signature);
         }
 
-        ensure_len(file_data, offset, 4 + 1 + 1 + 4 + 2 + 2 + 1 + 1)?;
-        let version = file_data[offset + 4];
+        w.ensure(0, 4 + 1 + 1 + 4 + 2 + 2 + 1 + 1)?;
+        let version = d[4];
         if version != 0 {
             return Err(FormatError::InvalidBTreeV2Version(version));
         }
 
-        let tree_type = file_data[offset + 5];
-        let node_size = u32::from_le_bytes([
-            file_data[offset + 6],
-            file_data[offset + 7],
-            file_data[offset + 8],
-            file_data[offset + 9],
-        ]);
-        let record_size = u16::from_le_bytes([file_data[offset + 10], file_data[offset + 11]]);
-        let depth = u16::from_le_bytes([file_data[offset + 12], file_data[offset + 13]]);
-        let _split_percent = file_data[offset + 14];
-        let _merge_percent = file_data[offset + 15];
+        let tree_type = d[5];
+        let node_size = u32::from_le_bytes([d[6], d[7], d[8], d[9]]);
+        let record_size = u16::from_le_bytes([d[10], d[11]]);
+        let depth = u16::from_le_bytes([d[12], d[13]]);
+        let _split_percent = d[14];
+        let _merge_percent = d[15];
 
-        let mut pos = offset + 16;
-        let root_node_address = read_offset(file_data, pos, offset_size)?;
+        let mut pos = 16;
+        w.ensure(pos, usize::from(offset_size))?;
+        let root_node_address = read_offset(d, pos, offset_size)?;
         pos += offset_size as usize;
 
-        ensure_len(file_data, pos, 2)?;
-        let num_records_in_root = u16::from_le_bytes([file_data[pos], file_data[pos + 1]]);
+        w.ensure(pos, 2)?;
+        let num_records_in_root = u16::from_le_bytes([d[pos], d[pos + 1]]);
         pos += 2;
 
-        let total_records = read_offset(file_data, pos, length_size)?;
+        w.ensure(pos, usize::from(length_size))?;
+        let total_records = read_offset(d, pos, length_size)?;
         #[allow(unused_assignments)]
         {
             pos += length_size as usize;
@@ -137,9 +154,9 @@ impl BTreeV2Header {
         // Validate header checksum
         #[cfg(feature = "checksum")]
         {
-            ensure_len(file_data, pos, 4)?;
-            let stored = LittleEndian::read_u32(&file_data[pos..pos + 4]);
-            let computed = crate::checksum::jenkins_lookup3(&file_data[offset..pos]);
+            w.ensure(pos, 4)?;
+            let stored = LittleEndian::read_u32(&d[pos..pos + 4]);
+            let computed = crate::checksum::jenkins_lookup3(&d[..pos]);
             if computed != stored {
                 return Err(FormatError::ChecksumMismatch {
                     expected: stored,
@@ -163,7 +180,7 @@ impl BTreeV2Header {
 /// Compute maximum records per node for a given depth level.
 /// leaf: (node_size - overhead) / record_size
 /// internal: depends on pointers
-fn max_records_leaf(node_size: u32, record_size: u16) -> u64 {
+pub(crate) fn max_records_leaf(node_size: u32, record_size: u16) -> u64 {
     // Leaf overhead: signature(4) + version(1) + type(1) + checksum(4) = 10
     let overhead = 10u32;
     if node_size <= overhead || record_size == 0 {
@@ -172,33 +189,79 @@ fn max_records_leaf(node_size: u32, record_size: u16) -> u64 {
     ((node_size - overhead) / record_size as u32) as u64
 }
 
+/// Deepest B-tree v2 accepted. See [`collect_btree_v2_records`].
+const MAX_DEPTH: u16 = 64;
+
+/// Take `n` records from the traversal's budget, or refuse the tree.
+fn spend(budget: &mut usize, n: usize) -> Result<(), FormatError> {
+    match budget.checked_sub(n) {
+        Some(left) => {
+            *budget = left;
+            Ok(())
+        }
+        None => {
+            // Spent: a walk that goes on after a failure stops here.
+            *budget = 0;
+            Err(FormatError::NestingDepthExceeded)
+        }
+    }
+}
+
 /// Collect all records from a B-tree v2 by traversing from the root.
 pub fn collect_btree_v2_records(
     file_data: &[u8],
     header: &BTreeV2Header,
     offset_size: u8,
     length_size: u8,
+) -> Result<Vec<BTreeV2Record>, FormatError> {
+    collect_btree_v2_records_in(file_data, header, offset_size, length_size)
+}
+
+/// [`collect_btree_v2_records`] over any [`Storage`]: one bounded read per
+/// node.
+pub fn collect_btree_v2_records_in<S: Storage + ?Sized>(
+    file: &S,
+    header: &BTreeV2Header,
+    offset_size: u8,
+    length_size: u8,
 ) -> Result<Vec<BTreeV2Record>, FormatError> {
     if header.total_records == 0 || header.num_records_in_root == 0 {
         return Ok(Vec::new());
     }
+    // Recursion is one frame per level, and the depth is read from the file:
+    // a crafted header claiming 65 535 levels over a node that is its own
+    // child overflowed the stack. 64 matches the fractal heap's guard, and no
+    // real tree comes close — even at the minimum fan-out of two it would
+    // hold more than 2^64 records.
+    if header.depth > MAX_DEPTH {
+        return Err(FormatError::NestingDepthExceeded);
+    }
+    // A valid tree stores each record once, in its own bytes, so it cannot
+    // hold more records than the file has room for. Children are addresses,
+    // though, and nothing makes them distinct: levels whose children all
+    // point at one shared node below reach it fan-out^depth times, which is
+    // millions of records from a few kilobytes. Counting against what the
+    // file could physically contain bounds that without trusting the
+    // header's own `total_records`.
+    let mut budget = len_usize(file) / usize::from(header.record_size.max(1));
 
     let max_leaf_nrec = max_records_leaf(header.node_size, header.record_size);
 
     if header.depth == 0 {
         // Root is a leaf
         parse_leaf_records(
-            file_data,
-            header.root_node_address as usize,
+            file,
+            to_usize(header.root_node_address)?,
             header.num_records_in_root,
             header.record_size,
+            header.node_size,
         )
     } else {
         // Root is internal; traverse recursively
         let mut records = Vec::new();
         collect_internal_records(
-            file_data,
-            header.root_node_address as usize,
+            file,
+            to_usize(header.root_node_address)?,
             header.num_records_in_root,
             header.depth,
             header.record_size,
@@ -206,42 +269,79 @@ pub fn collect_btree_v2_records(
             offset_size,
             length_size,
             max_leaf_nrec,
+            &mut budget,
             &mut records,
         )?;
         Ok(records)
     }
 }
 
+/// A node's bytes: `want` bytes at `offset` (fewer only at the end of the
+/// file), after checking its 4-byte signature. A node is read in one piece
+/// when it fits in `node_size` (every valid node does); a larger claimed
+/// extent — record counts from a damaged parent — is first checked against
+/// the end of the file, so it costs a read only of bytes the file has.
+/// Bounds errors are the whole-file ones: the signature check needs the
+/// first 6 bytes, then `checks` — `(position, length)` pairs relative to
+/// the node, in the order the parser checks them — must lie in the file.
+fn read_node<'a, S: Storage + ?Sized>(
+    file: &'a S,
+    offset: usize,
+    want: usize,
+    node_size: u32,
+    signature: &[u8; 4],
+    checks: &[(usize, usize)],
+) -> Result<Window<'a>, FormatError> {
+    let one_read = usize::try_from(node_size).unwrap_or(usize::MAX).max(6);
+    let w = Window::read(file, offset as u64, want.min(one_read))?;
+    w.ensure(0, 6)?;
+    if &w.bytes[..4] != signature {
+        return Err(FormatError::InvalidBTreeV2Signature);
+    }
+    if want <= one_read {
+        return Ok(w);
+    }
+    for &(rel, len) in checks {
+        Window::check_extent(file, offset as u64, rel, len)?;
+    }
+    Window::read(file, offset as u64, want)
+}
+
 /// Parse records from a leaf node (signature "BTLF").
-fn parse_leaf_records(
-    file_data: &[u8],
+fn parse_leaf_records<S: Storage + ?Sized>(
+    file: &S,
     offset: usize,
     num_records: u16,
     record_size: u16,
+    node_size: u32,
 ) -> Result<Vec<BTreeV2Record>, FormatError> {
     // signature(4) + version(1) + type(1) = 6 bytes header
-    ensure_len(file_data, offset, 6)?;
-    if &file_data[offset..offset + 4] != b"BTLF" {
-        return Err(FormatError::InvalidBTreeV2Signature);
-    }
-
-    let pos = offset + 6;
+    let pos = 6;
     let rs = record_size as usize;
     let total = (num_records as usize)
         .checked_mul(rs)
         .ok_or(FormatError::UnexpectedEof {
             expected: usize::MAX,
-            available: file_data.len(),
+            available: len_usize(file),
         })?;
-    ensure_len(file_data, pos, total)?;
+    let w = read_node(
+        file,
+        offset,
+        pos + total + 4,
+        node_size,
+        b"BTLF",
+        &[(pos, total)],
+    )?;
+    let d = &w.bytes;
+    w.ensure(pos, total)?;
 
     // Validate checksum: 4 bytes after records + padding
     #[cfg(feature = "checksum")]
     {
         let checksum_pos = pos + total;
-        if file_data.len() >= checksum_pos + 4 {
-            let stored = LittleEndian::read_u32(&file_data[checksum_pos..checksum_pos + 4]);
-            let computed = crate::checksum::jenkins_lookup3(&file_data[offset..checksum_pos]);
+        if d.len() >= checksum_pos + 4 {
+            let stored = LittleEndian::read_u32(&d[checksum_pos..checksum_pos + 4]);
+            let computed = crate::checksum::jenkins_lookup3(&d[..checksum_pos]);
             if computed != stored {
                 return Err(FormatError::ChecksumMismatch {
                     expected: stored,
@@ -255,16 +355,135 @@ fn parse_leaf_records(
     for i in 0..num_records as usize {
         let start = pos + i * rs;
         records.push(BTreeV2Record {
-            data: file_data[start..start + rs].to_vec(),
+            data: d[start..start + rs].to_vec(),
         });
     }
     Ok(records)
 }
 
+/// An internal node read from the file: its bytes (from the signature on),
+/// where its records start, and its children as `(address, record count)`.
+struct InternalNode<'a> {
+    node: Window<'a>,
+    records_start: usize,
+    children: Vec<(u64, u16)>,
+}
+
+impl InternalNode<'_> {
+    /// Record `i`, `rs` bytes long.
+    fn record(&self, i: usize, rs: usize) -> Result<&[u8], FormatError> {
+        let overflow = || FormatError::UnexpectedEof {
+            expected: usize::MAX,
+            available: usize::MAX,
+        };
+        let rec_start = i
+            .checked_mul(rs)
+            .and_then(|o| self.records_start.checked_add(o))
+            .ok_or_else(overflow)?;
+        self.node.ensure(rec_start, rs)?;
+        Ok(&self.node.bytes[rec_start..rec_start + rs])
+    }
+}
+
+/// An internal node's layout: where its records start, and its children as
+/// `(address, record count)`.
+#[allow(clippy::too_many_arguments)]
+fn read_internal_node<S: Storage + ?Sized>(
+    file: &S,
+    offset: usize,
+    num_records: u16,
+    depth: u16,
+    record_size: u16,
+    node_size: u32,
+    offset_size: u8,
+    max_leaf_nrec: u64,
+) -> Result<InternalNode<'_>, FormatError> {
+    let nr = num_records as usize;
+    let rs = record_size as usize;
+
+    // Records first
+    let records_total = nr.checked_mul(rs).ok_or(FormatError::UnexpectedEof {
+        expected: usize::MAX,
+        available: len_usize(file),
+    })?;
+
+    // Child pointer layout, as libhdf5 computes it (H5B2__hdr_init): the
+    // child's record count is always encoded in the width needed for a
+    // *leaf's* maximum, and — below the first internal level — the child
+    // subtree's total record count in the width needed for the most records
+    // a subtree of that depth can hold.
+    let child_depth = depth - 1;
+    let nrec_width = bytes_for_max_records(max_leaf_nrec);
+    let total_nrec_width = if depth > 1 {
+        bytes_for_max_records(cum_max_records(
+            node_size,
+            record_size,
+            offset_size,
+            max_leaf_nrec,
+            child_depth,
+        ))
+    } else {
+        0
+    };
+
+    let num_children = nr + 1;
+    let child_ptr_size = offset_size as usize + nrec_width + total_nrec_width;
+    let pointers = num_children * child_ptr_size;
+
+    // signature(4) + version(1) + type(1) = 6, records, pointers, checksum.
+    let w = read_node(
+        file,
+        offset,
+        6 + records_total + pointers + 4,
+        node_size,
+        b"BTIN",
+        &[(6, records_total), (6 + records_total, pointers)],
+    )?;
+    let d = &w.bytes;
+    let mut pos = 6;
+    w.ensure(pos, records_total)?;
+    let records_start = pos;
+    pos += records_total;
+
+    w.ensure(pos, pointers)?;
+
+    let mut children = Vec::with_capacity(num_children);
+    for _ in 0..num_children {
+        let addr = read_offset(d, pos, offset_size)?;
+        pos += offset_size as usize;
+        let child_nrec = read_var_uint(d, pos, nrec_width)? as u16;
+        pos += nrec_width;
+        pos += total_nrec_width; // skip total records in subtree
+        children.push((addr, child_nrec));
+    }
+
+    // The checksum follows the child pointers and covers the node up to it.
+    // Lookups prune children by the keys in this node, so an unverified
+    // internal node could hide a record without any error: libhdf5 refuses
+    // a mismatch here, and so does this.
+    #[cfg(feature = "checksum")]
+    {
+        w.ensure(pos, 4)?;
+        let stored = LittleEndian::read_u32(&d[pos..pos + 4]);
+        let computed = crate::checksum::jenkins_lookup3(&d[..pos]);
+        if computed != stored {
+            return Err(FormatError::ChecksumMismatch {
+                expected: stored,
+                computed,
+            });
+        }
+    }
+    Ok(InternalNode {
+        node: w,
+        records_start,
+        children,
+    })
+}
+
 /// Recursively collect records from an internal node.
 #[allow(clippy::too_many_arguments, clippy::only_used_in_recursion)]
-fn collect_internal_records(
-    file_data: &[u8],
+fn collect_internal_records<S: Storage + ?Sized>(
+    file: &S,
     offset: usize,
     num_records: u16,
     depth: u16,
@@ -273,90 +492,54 @@ fn collect_internal_records(
     offset_size: u8,
     length_size: u8,
     max_leaf_nrec: u64,
+    budget: &mut usize,
     out: &mut Vec<BTreeV2Record>,
 ) -> Result<(), FormatError> {
-    // signature(4) + version(1) + type(1) = 6
-    ensure_len(file_data, offset, 6)?;
-    if &file_data[offset..offset + 4] != b"BTIN" {
-        return Err(FormatError::InvalidBTreeV2Signature);
-    }
-
     let nr = num_records as usize;
     let rs = record_size as usize;
-    let mut pos = offset + 6;
-
-    // Read all records first
-    let records_total = nr.checked_mul(rs).ok_or(FormatError::UnexpectedEof {
-        expected: usize::MAX,
-        available: file_data.len(),
-    })?;
-    ensure_len(file_data, pos, records_total)?;
-    let records_start = pos;
-    pos += records_total;
-
-    // Compute sizes for child pointers
-    // max_records at child depth - for variable-width nrec encoding
+    let node = read_internal_node(
+        file,
+        offset,
+        num_records,
+        depth,
+        record_size,
+        node_size,
+        offset_size,
+        max_leaf_nrec,
+    )?;
     let child_depth = depth - 1;
-    let max_nrec_child = if child_depth == 0 {
-        max_leaf_nrec
-    } else {
-        // For internal nodes at child_depth, the true max_nrec depends on the
-        // node size, record size, and the recursive width of child pointer
-        // entries (which themselves depend on max_nrec at deeper levels).
-        // Computing the exact value requires iterating from the leaf level
-        // upward, as described in the HDF5 spec (III.A.2 "Computing the Size
-        // of B-tree Nodes").
-        //
-        // We use `max_leaf_nrec * 2` as a conservative upper bound. This
-        // over-estimates the nrec encoding width, which means we may read
-        // slightly more bytes per child pointer than strictly necessary, but
-        // never fewer. The over-read bytes are harmless because we only
-        // decode `num_records` entries (the actual count from the node header).
-        //
-        // Known limitation: for very deep trees (depth > 3) with small record
-        // sizes, the true max could exceed this estimate, causing us to
-        // under-allocate the nrec encoding width and misparse child pointers.
-        // In practice, HDF5 B-tree v2 depths rarely exceed 2-3.
-        max_leaf_nrec * 2
-    };
-    let nrec_width = bytes_for_max_records(max_nrec_child);
-
-    // Total records in subtree width (only if depth > 1)
-    let total_nrec_width = if depth > 1 {
-        // Width to hold total records in a subtree
-        // We compute max possible total records at this subtree depth
-        let max_total = header_max_total_records(max_leaf_nrec, depth - 1);
-        bytes_for_max_records(max_total)
-    } else {
-        0
-    };
-
-    let num_children = nr + 1;
-    let child_ptr_size = offset_size as usize + nrec_width + total_nrec_width;
-    ensure_len(file_data, pos, num_children * child_ptr_size)?;
-
-    // Read child pointers
-    let mut children = Vec::with_capacity(num_children);
-    for _ in 0..num_children {
-        let addr = read_offset(file_data, pos, offset_size)?;
-        pos += offset_size as usize;
-        let child_nrec = read_var_uint(file_data, pos, nrec_width)? as u16;
-        pos += nrec_width;
-        pos += total_nrec_width; // skip total records in subtree
-        children.push((addr, child_nrec));
-    }
 
     // Interleave: child[0], record[0], child[1], record[1], ..., child[nr]
     // We collect child[0] records, then record[0], then child[1], etc.
-    for (i, &(child_addr, child_nrec)) in children.iter().enumerate() {
+    // A child that fails does not stop the walk: the others are still
+    // descended into (their records are dropped with the result), then the
+    // first error is returned, as when stopping there. A storage that
+    // records what it lacks (see `storage::touch`) so learns every node the
+    // walk can reach in one attempt. The record budget is spent as before,
+    // so the walk is no longer than a successful one.
+    let mut failed = None;
+    for (i, &(child_addr, child_nrec)) in node.children.iter().enumerate() {
+        if failed.is_some() && *budget == 0 {
+            // The record budget is spent: the tree is refused, and a walk
+            // over what is left could be as long as the one it bounds.
+            break;
+        }
+        if let Err(e) = (|| -> Result<(), FormatError> {
             if child_depth == 0 {
-            let leaf_recs =
-                parse_leaf_records(file_data, child_addr as usize, child_nrec, record_size)?;
+                // Before parsing, so a refused tree is not also a large allocation.
+                spend(budget, usize::from(child_nrec))?;
+                let leaf_recs = parse_leaf_records(
+                    file,
+                    to_usize(child_addr)?,
+                    child_nrec,
+                    record_size,
+                    node_size,
+                )?;
                 out.extend(leaf_recs);
             } else {
                 collect_internal_records(
-                file_data,
-                child_addr as usize,
+                    file,
+                    to_usize(child_addr)?,
                     child_nrec,
                     child_depth,
                     record_size,
@@ -364,52 +547,243 @@ fn collect_internal_records(
                     offset_size,
                     length_size,
                     max_leaf_nrec,
+                    budget,
                     out,
                 )?;
             }
 
             // Add record[i] (except after the last child)
             if i < nr {
-            let rec_offset = i.checked_mul(rs).ok_or(FormatError::UnexpectedEof {
-                expected: usize::MAX,
-                available: file_data.len(),
-            })?;
-            let rec_start =
-                records_start
-                    .checked_add(rec_offset)
-                    .ok_or(FormatError::UnexpectedEof {
-                        expected: usize::MAX,
-                        available: file_data.len(),
-                    })?;
-            let rec_end = rec_start
-                .checked_add(rs)
-                .ok_or(FormatError::UnexpectedEof {
-                    expected: usize::MAX,
-                    available: file_data.len(),
-                })?;
-            if rec_end > file_data.len() {
-                return Err(FormatError::UnexpectedEof {
-                    expected: rec_end,
-                    available: file_data.len(),
+                let data = node.record(i, rs)?;
+                spend(budget, 1)?;
+                out.push(BTreeV2Record {
+                    data: data.to_vec(),
                 });
             }
-            out.push(BTreeV2Record {
-                data: file_data[rec_start..rec_end].to_vec(),
-            });
+            Ok(())
+        })() {
+            failed.get_or_insert(e);
         }
     }
 
+    match failed {
+        Some(e) => Err(e),
+        None => Ok(()),
+    }
+}
+
+/// The records of a B-tree v2 that fall in one key range, found by
+/// descending the tree instead of reading all of it.
+///
+/// `cmp` places a record relative to the range: `Less` if the record sorts
+/// before it, `Greater` if after, `Equal` if the record is in it. The tree
+/// must be ordered consistently with `cmp`, as libhdf5 orders it (a link or
+/// attribute name index by name hash, so all records with one hash form a
+/// range whatever order their names are in). Only the nodes whose key
+/// interval overlaps the range are read: O(depth) nodes plus those holding
+/// the matches. Matches come in tree order.
+pub fn find_btree_v2_records(
+    file_data: &[u8],
+    header: &BTreeV2Header,
+    offset_size: u8,
+    cmp: &mut dyn FnMut(&[u8]) -> Ordering,
+) -> Result<Vec<BTreeV2Record>, FormatError> {
+    find_btree_v2_records_in(file_data, header, offset_size, cmp)
+}
+
+/// [`find_btree_v2_records`] over any [`Storage`]: one bounded read per
+/// node visited.
+pub fn find_btree_v2_records_in<S: Storage + ?Sized>(
+    file: &S,
+    header: &BTreeV2Header,
+    offset_size: u8,
+    cmp: &mut dyn FnMut(&[u8]) -> Ordering,
+) -> Result<Vec<BTreeV2Record>, FormatError> {
+    if header.total_records == 0 || header.num_records_in_root == 0 {
+        return Ok(Vec::new());
+    }
+    if header.depth > MAX_DEPTH {
+        return Err(FormatError::NestingDepthExceeded);
+    }
+    // As in `collect_btree_v2_records`: a valid tree cannot hold more
+    // records than the file has room for, however its children are shared.
+    let mut budget = len_usize(file) / usize::from(header.record_size.max(1));
+    let max_leaf_nrec = max_records_leaf(header.node_size, header.record_size);
+    let mut out = Vec::new();
+    find_in_node(
+        file,
+        header,
+        to_usize(header.root_node_address)?,
+        header.num_records_in_root,
+        header.depth,
+        offset_size,
+        max_leaf_nrec,
+        cmp,
+        &mut budget,
+        &mut out,
+    )?;
+    Ok(out)
+}
+
+#[allow(clippy::too_many_arguments)]
+fn find_in_node<S: Storage + ?Sized>(
+    file: &S,
+    header: &BTreeV2Header,
+    offset: usize,
+    num_records: u16,
+    depth: u16,
+    offset_size: u8,
+    max_leaf_nrec: u64,
+    cmp: &mut dyn FnMut(&[u8]) -> Ordering,
+    budget: &mut usize,
+    out: &mut Vec<BTreeV2Record>,
+) -> Result<(), FormatError> {
+    spend(budget, usize::from(num_records))?;
+    if depth == 0 {
+        let records = parse_leaf_records(
+            file,
+            offset,
+            num_records,
+            header.record_size,
+            header.node_size,
+        )?;
+        out.extend(
+            records
+                .into_iter()
+                .filter(|r| cmp(&r.data) == Ordering::Equal),
+        );
+        return Ok(());
+    }
+    let rs = usize::from(header.record_size);
+    let node = read_internal_node(
+        file,
+        offset,
+        num_records,
+        depth,
+        header.record_size,
+        header.node_size,
+        offset_size,
+        max_leaf_nrec,
+    )?;
+    let nr = usize::from(num_records);
+    let mut order = Vec::with_capacity(nr);
+    for i in 0..nr {
+        order.push(cmp(node.record(i, rs)?));
+    }
+    // Child `i` holds the keys between record `i - 1` and record `i`: it can
+    // hold a match unless the record before it is already past the range or
+    // the record after it is still before it.
+    for (i, &(child_addr, child_nrec)) in node.children.iter().enumerate() {
+        let after_left = i == 0 || order[i - 1] != Ordering::Greater;
+        let before_right = i == nr || order[i] != Ordering::Less;
+        if after_left && before_right {
+            find_in_node(
+                file,
+                header,
+                to_usize(child_addr)?,
+                child_nrec,
+                depth - 1,
+                offset_size,
+                max_leaf_nrec,
+                cmp,
+                budget,
+                out,
+            )?;
+        }
+        if i < nr && order[i] == Ordering::Equal {
+            out.push(BTreeV2Record {
+                data: node.record(i, rs)?.to_vec(),
+            });
+        }
+    }
     Ok(())
 }
 
-/// Estimate maximum total records at a given depth (for variable-width encoding).
-fn header_max_total_records(max_leaf_nrec: u64, depth: u16) -> u64 {
-    // Conservative: branching factor * max_leaf at each level
-    let mut total = max_leaf_nrec;
-    for _ in 0..depth {
-        total = total.saturating_mul(max_leaf_nrec.max(2));
+/// Most records a subtree whose root is at `depth` can hold (libhdf5's
+/// `cum_max_nrec`). See [`node_info`].
+fn cum_max_records(
+    node_size: u32,
+    record_size: u16,
+    offset_size: u8,
+    max_leaf_nrec: u64,
+    depth: u16,
+) -> u64 {
+    node_info_from_leaf(node_size, record_size, offset_size, max_leaf_nrec, depth)
+        .last()
+        .map_or(max_leaf_nrec, |n| n.cum_max_nrec)
+}
+
+/// Capacity of a B-tree v2 node at one depth, as libhdf5 computes it
+/// (`H5B2__hdr_init`'s `node_info`).
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub(crate) struct NodeInfo {
+    /// Most records one node at this depth holds.
+    pub(crate) max_nrec: u64,
+    /// Most records a subtree rooted at this depth holds.
+    pub(crate) cum_max_nrec: u64,
+    /// Bytes a subtree's total record count takes in a pointer to a node
+    /// at this depth (0 for a leaf, whose count is its own).
+    pub(crate) cum_max_nrec_size: usize,
+}
+
+/// Node capacities for depths `0..=depth` (entry `d` for depth `d`): a leaf
+/// holds `max_nrec(0)` records; an internal node at depth `d` holds
+/// `max_nrec(d)` records and `max_nrec(d) + 1` subtrees of depth `d - 1`,
+/// where `max_nrec(d)` is what fits in a node once each record is paired
+/// with a child pointer of the width depth `d` needs (address, the child's
+/// record count in the width a *leaf's* maximum needs, and below the first
+/// internal level the child subtree's total in the width its maximum
+/// needs), with one pointer more than records.
+pub(crate) fn node_info(
+    node_size: u32,
+    record_size: u16,
+    offset_size: u8,
+    depth: u16,
+) -> Vec<NodeInfo> {
+    let max_leaf = max_records_leaf(node_size, record_size);
+    node_info_from_leaf(node_size, record_size, offset_size, max_leaf, depth)
+}
+
+fn node_info_from_leaf(
+    node_size: u32,
+    record_size: u16,
+    offset_size: u8,
+    max_leaf_nrec: u64,
+    depth: u16,
+) -> Vec<NodeInfo> {
+    // Internal node overhead: signature(4) + version(1) + type(1) + checksum(4).
+    const PREFIX: u64 = 10;
+    let nrec_width = bytes_for_max_records(max_leaf_nrec) as u64;
+    let mut info = Vec::with_capacity(usize::from(depth) + 1);
+    info.push(NodeInfo {
+        max_nrec: max_leaf_nrec,
+        cum_max_nrec: max_leaf_nrec,
+        cum_max_nrec_size: 0,
+    });
+    for d in 1..=depth {
+        let below = info[usize::from(d) - 1];
+        let ptr = u64::from(offset_size)
+            + nrec_width
+            + if d > 1 {
+                below.cum_max_nrec_size as u64
+            } else {
+                0
+            };
+        let max_nrec = u64::from(node_size)
+            .saturating_sub(PREFIX)
+            .saturating_sub(ptr)
+            / (u64::from(record_size) + ptr).max(1);
+        let cum = max_nrec
+            .saturating_add(1)
+            .saturating_mul(below.cum_max_nrec)
+            .saturating_add(max_nrec);
+        info.push(NodeInfo {
+            max_nrec,
+            cum_max_nrec: cum,
+            cum_max_nrec_size: bytes_for_max_records(cum),
+        });
     }
-    total
+    info
 }
 
 #[cfg(test)]
@@ -466,6 +840,132 @@ mod tests {
         buf
     }
 
+    /// An internal node laid out exactly as `collect_internal_records` will
+    /// read it at `depth`: `records` zeroed records, then `children` pointers,
+    /// all to `child_addr` claiming `child_nrec` records.
+    fn internal_node(
+        depth: u16,
+        node_size: u32,
+        record_size: u16,
+        records: usize,
+        children: usize,
+        child_addr: u64,
+        child_nrec: u64,
+    ) -> Vec<u8> {
+        let max_leaf = max_records_leaf(node_size, record_size);
+        let nrec_width = bytes_for_max_records(max_leaf);
+        let total_width = if depth > 1 {
+            bytes_for_max_records(cum_max_records(
+                node_size,
+                record_size,
+                8,
+                max_leaf,
+                depth - 1,
+            ))
+        } else {
+            0
+        };
+        let mut buf = b"BTIN".to_vec();
+        buf.extend_from_slice(&[0, 5]);
+        buf.resize(buf.len() + records * record_size as usize, 0);
+        for _ in 0..children {
+            buf.extend_from_slice(&child_addr.to_le_bytes());
+            buf.extend_from_slice(&child_nrec.to_le_bytes()[..nrec_width]);
+            buf.resize(buf.len() + total_width, 0);
+        }
+        let sum = crate::checksum::jenkins_lookup3(&buf);
+        buf.extend_from_slice(&sum.to_le_bytes());
+        buf
+    }
+
+    fn header(depth: u16, root: u64, root_nrec: u16, total: u64) -> BTreeV2Header {
+        BTreeV2Header {
+            tree_type: 5,
+            node_size: 512,
+            record_size: 8,
+            depth,
+            root_node_address: root,
+            num_records_in_root: root_nrec,
+            total_records: total,
+        }
+    }
+
+    #[test]
+    fn a_node_that_is_its_own_child_is_rejected_not_recursed() {
+        // One internal node whose two children are itself, under a header
+        // claiming the deepest tree a u16 allows. The layout stops depending
+        // on depth once the subtree-total width saturates, so every level
+        // parses cleanly and recursion runs ~65 000 frames deep: before the
+        // cap this overflowed the stack and aborted the process, from a file
+        // of under 100 bytes.
+        let mut data = internal_node(u16::MAX, 512, 8, 1, 2, 0, 1);
+        data.resize(4096, 0);
+        let result = collect_btree_v2_records(&data, &header(u16::MAX, 0, 1, 1), 8, 8);
+        assert!(result.is_err(), "{result:?}");
+    }
+
+    #[test]
+    fn a_shared_subtree_cannot_multiply_the_work() {
+        // A chain of distinct levels, each node's children all pointing at the
+        // single node below, ending in a real leaf. Every node parses and
+        // nothing is cyclic, yet the leaf is reached fan-out^depth times: 62
+        // children over 4 levels is ~15 million leaf visits from a few
+        // kilobytes. A valid tree cannot hold more records than the file has
+        // room for, so that bounds the traversal instead.
+        let (node_size, record_size) = (512u32, 8u16);
+        let fanout = 62usize;
+        let depth = 4u16;
+        let leaf = build_leaf_node(5, &[&[0u8; 8][..]]);
+
+        // Lay out root first, then each lower level, then the leaf.
+        let mut nodes: Vec<Vec<u8>> = Vec::new();
+        let mut addrs = Vec::new();
+        let mut at = 0u64;
+        let mut sizes = Vec::new();
+        for d in (1..=depth).rev() {
+            let n = internal_node(d, node_size, record_size, fanout - 1, fanout, 0, 0);
+            sizes.push(n.len());
+        }
+        for size in &sizes {
+            addrs.push(at);
+            at += *size as u64;
+        }
+        let leaf_addr = at;
+        for (i, d) in (1..=depth).rev().enumerate() {
+            let (child, child_nrec) = if d == 1 {
+                (leaf_addr, 1)
+            } else {
+                (addrs[i + 1], fanout as u64 - 1)
+            };
+            nodes.push(internal_node(
+                d,
+                node_size,
+                record_size,
+                fanout - 1,
+                fanout,
+                child,
+                child_nrec,
+            ));
+        }
+        let mut data: Vec<u8> = nodes.concat();
+        data.extend_from_slice(&leaf);
+        data.resize(data.len() + 64, 0);
+
+        let started = std::time::Instant::now();
+        let result =
+            collect_btree_v2_records(&data, &header(depth, 0, fanout as u16 - 1, u64::MAX), 8, 8);
+        assert!(
+            result.is_err(),
+            "expected a refusal, got {} records",
+            result.map_or(0, |r| r.len())
+        );
+        assert!(
+            started.elapsed() < std::time::Duration::from_secs(2),
+            "took {:?}",
+            started.elapsed()
+        );
+    }
+
     #[test]
     fn parse_header() {
         let data = build_btree_v2_header(5, 512, 11, 0, 0x1000, 3, 3, 8, 8);
@@ -522,4 +1022,18 @@ mod tests {
         let records = collect_btree_v2_records(&header, &hdr, 8, 8).unwrap();
         assert!(records.is_empty());
     }
+
+    #[test]
+    fn subtree_capacity_matches_libhdf5() {
+        // A link-name index (11-byte records, 512-byte nodes, 8-byte
+        // addresses): libhdf5's H5B2__hdr_init gives 45 records per leaf,
+        // then cum_max_nrec 1 149 at depth 1 and 26 449 at depth 2 — two
+        // bytes of subtree count in a depth-3 root's child pointers, where
+        // leaf_max^3 = 91 125 would need three.
+        let leaf = max_records_leaf(512, 11);
+        assert_eq!(leaf, 45);
+        assert_eq!(cum_max_records(512, 11, 8, leaf, 0), 45);
+        assert_eq!(cum_max_records(512, 11, 8, leaf, 1), 1_149);
+        assert_eq!(cum_max_records(512, 11, 8, leaf, 2), 26_449);
+    }
 }