Merge release/v2.6.0

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
chore(release): v2.6.0
2026-09-20 06:01:38 -07:00 · 2026-09-20 06:01:31 -07:00 · 2026-09-20 04:55:09 -07:00 · 2026-09-19 20:45:21 -07:00 · 2026-09-19 20:44:59 -07:00 · 2026-09-19 20:44:54 -07:00
132 changed files with 13738 additions and 2149 deletions
@@ -22,5 +22,25 @@ jobs:
        run: rustup component add rustfmt clippy
      - name: Install thumbv7em-none-eabihf target
        run: rustup target add thumbv7em-none-eabihf
+      - 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
+          python3 -m venv /opt/interop
+          /opt/interop/bin/pip install --no-cache-dir h5py numpy netCDF4 xarray
+          echo "/opt/interop/bin" >> "$GITHUB_PATH"
+      - name: Show interop library versions
+        run: /opt/interop/bin/python -c "import h5py, netCDF4; print('h5py', h5py.__version__, 'HDF5', h5py.version.hdf5_version, 'netCDF4', netCDF4.__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
@@ -4,3 +4,4 @@ benchmarks/longmemeval/*.json

 # Local model weights (MiniLM etc.) — large, not committed
 weights/
+.venv
@@ -28,6 +28,402 @@

 ---

+## Memory footprint
+
+`cargo run --release -p clawhdf5-bench --bin search_harness -- --footprint --full`,
+384-dim `f32`. The figure that matters is **reopened**: a store loaded from
+disk, which is what a long-lived process holds.
+
+Measured with a counting global allocator, not RSS. RSS cannot see 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 at all. Measured that way a store holding the corpus twice and one
+holding it once came out *identical* (1.00x both), which is how the first
+attempt at this measurement went.
+
+| N | vectors (raw) | reopened, before | reopened, after |
+|---:|---:|---:|---:|
+| 1 000 | 1 MiB | 5 MiB (3.41x) | 4 MiB (2.39x) |
+| 10 000 | 15 MiB | 50 MiB (3.43x) | 35 MiB (2.42x) |
+| 100 000 | 146 MiB | 505 MiB (3.44x) | **357 MiB (2.43x)** |
+
+The cache stored every embedding twice — once as a `Vec<Vec<f32>>` and once
+flattened for the batched kernels, kept in lock-step on every push, update and
+compaction. Storing only the flat buffer and indexing into it gives back
+almost exactly one copy of the corpus (148 MiB at 100k) and one heap
+allocation per entry. Recall and query latency are unchanged.
+
+What remains at 2.43x: the flat vectors (1.0x), the HNSW index's own copy of
+them (1.0x), and text, ids and graph (~0.4x). The index copy is the next
+target — it is what a quantised or borrowed representation would address.
+
+### Quantising the index copy (`quantized_index`)
+
+`MemoryConfig::quantized_index` stores the index's copy as `i8` instead of
+`f32`. Same harness, same binary, `--footprint --full` with and without
+`--int8`:
+
+| N | vectors (raw) | indexes, f32 | indexes, int8 | reopened, f32 | reopened, int8 |
+|---:|---:|---:|---:|---:|---:|
+| 1 000 | 1 MiB | 2 MiB | 1 MiB | 4 MiB (2.40x) | 2 MiB (1.64x) |
+| 10 000 | 15 MiB | 32 MiB | 14 MiB | 44 MiB (3.03x) | 27 MiB (1.81x) |
+| 100 000 | 146 MiB | 266 MiB | **123 MiB** | 399 MiB (2.72x) | **256 MiB (1.74x)** |
+
+The scale is **per row**, not 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: measured against an exact
+ranking that gives 0.35 top-10 overlap — unusable. Scaling each row by its own
+largest component brings the same measurement to 0.99.
+
+Quantised distances still cost recall on their own, and **`ef` does not buy it
+back**, because the loss is in the distances rather than in the graph
+(`--ann-only --full`, N = 100 000):
+
+| ef | recall@10, f32 | recall@10, int8 | recall@10, int8 + re-score |
+|---:|---:|---:|---:|
+| 32 | 0.9775 | 0.9415 | 0.9785 |
+| 64 | 0.9945 | 0.9625 | 0.9940 |
+| 128 | 0.9995 | 0.9670 | 0.9990 |
+| 256 | 0.9995 | 0.9670 (ceiling) | 0.9990 |
+
+Re-scoring closes the gap: the store already holds the exact embeddings, so
+the query path re-scores the candidate pool against them before fusion. That
+is done automatically whenever the index is quantised. What it costs is
+throughput — about 13% of QPS and 16% of build time at 100 000 x 384. So the
+setting trades ~13% of query speed for ~36% of the process's memory at equal
+recall. It is **off by default**: the right side of that trade depends on
+whether the deployment is short of memory or short of CPU.
+
+A measurement trap worth recording: the synthetic `clustered` generator in the
+`clawhdf5-ann` tests draws clusters far tighter than any real embedding, so
+neighbours there sit closer together than the quantisation error and top-10
+*identity* is noise. Scored on that fixture int8 looks catastrophic (0.57
+overlap) — a fact about the fixture, not the storage. The tests use random
+vectors, and recall is measured against brute-force ground truth rather than
+against the f32 index, whose own approximation errors a re-scored search is
+entitled to get right.
+
+## Read harness
+
+Produced by `cargo run --release -p clawhdf5-bench --bin read_harness`: a 4096 x
+2048 `f64` dataset (64 MB) written three ways, read in full and through four
+hyperslab selections, each from a fresh file handle. The last column is the
+point: does a selection cost what the *selection* costs?
+
+### Baseline (v2.4.0): every selection decodes the whole dataset
+
+4096 x 2048 f64 (64 MB per dataset), chunks 256 x 256, file 129 MB
+
+| layout | read | selected | time ms | MB/s of selection | vs full read |
+|---|---|---:|---:|---:|---:|
+| chunked + deflate | full (first) | 64 MB | 181.8 | 352 | |
+| chunked + deflate | full (repeat) | 64 MB | 162.1 | 395 | 1.00x |
+| chunked + deflate | 64 x 64 window (1 chunk) | 0.03 MB | 104.89 | 0 | 0.577x |
+| chunked + deflate | 512 x 512 window (4-9 chunks) | 2.00 MB | 110.26 | 18 | 0.606x |
+| chunked + deflate | one row | 0.02 MB | 105.81 | 0 | 0.582x |
+| chunked + deflate | one column | 0.03 MB | 108.37 | 0 | 0.596x |
+| chunked | full (first) | 64 MB | 97.4 | 657 | |
+| chunked | full (repeat) | 64 MB | 86.7 | 738 | 1.00x |
+| chunked | 64 x 64 window (1 chunk) | 0.03 MB | 40.97 | 1 | 0.420x |
+| chunked | 512 x 512 window (4-9 chunks) | 2.00 MB | 44.66 | 45 | 0.458x |
+| chunked | one row | 0.02 MB | 30.88 | 1 | 0.317x |
+| chunked | one column | 0.03 MB | 30.27 | 1 | 0.311x |
+| contiguous | full (first) | 64 MB | 57.6 | 1112 | |
+| contiguous | full (repeat) | 64 MB | 53.5 | 1195 | 1.00x |
+| contiguous | 64 x 64 window (1 chunk) | 0.03 MB | 30.97 | 1 | 0.538x |
+| contiguous | 512 x 512 window (4-9 chunks) | 2.00 MB | 31.64 | 63 | 0.550x |
+| contiguous | one row | 0.02 MB | 31.90 | 0 | 0.554x |
+| contiguous | one column | 0.03 MB | 29.36 | 1 | 0.510x |
+
+### After: partial reads
+
+Only the rows of a contiguous dataset, or the chunks, that overlap the
+selection's bounding box are read/decoded. A 64 x 64 window of the compressed
+dataset: **105 -> 0.39 ms**; one row: **106 -> 2.7 ms**; one column:
+**108 -> 5.2 ms**. (Absolute full-read times differ between the two runs
+because the machine's speed drifted; compare the *vs full read* column.)
+
+4096 x 2048 f64 (64 MB per dataset), chunks 256 x 256, file 129 MB
+
+| layout | read | selected | time ms | MB/s of selection | vs full read |
+|---|---|---:|---:|---:|---:|
+| chunked + deflate | full (first) | 64 MB | 112.5 | 569 | |
+| chunked + deflate | full (repeat) | 64 MB | 104.5 | 612 | 1.00x |
+| chunked + deflate | 64 x 64 window (1 chunk) | 0.03 MB | 0.39 | 81 | 0.003x |
+| chunked + deflate | 512 x 512 window (4-9 chunks) | 2.00 MB | 4.85 | 412 | 0.043x |
+| chunked + deflate | one row | 0.02 MB | 2.69 | 6 | 0.024x |
+| chunked + deflate | one column | 0.03 MB | 5.23 | 6 | 0.046x |
+| chunked | full (first) | 64 MB | 70.0 | 915 | |
+| chunked | full (repeat) | 64 MB | 61.7 | 1037 | 1.00x |
+| chunked | 64 x 64 window (1 chunk) | 0.03 MB | 0.06 | 541 | 0.001x |
+| chunked | 512 x 512 window (4-9 chunks) | 2.00 MB | 1.99 | 1005 | 0.028x |
+| chunked | one row | 0.02 MB | 0.05 | 285 | 0.001x |
+| chunked | one column | 0.03 MB | 0.45 | 69 | 0.006x |
+| contiguous | full (first) | 64 MB | 60.3 | 1062 | |
+| contiguous | full (repeat) | 64 MB | 56.4 | 1134 | 1.00x |
+| contiguous | 64 x 64 window (1 chunk) | 0.03 MB | 0.08 | 396 | 0.001x |
+| contiguous | 512 x 512 window (4-9 chunks) | 2.00 MB | 2.12 | 944 | 0.035x |
+| contiguous | one row | 0.02 MB | 0.03 | 576 | 0.000x |
+| contiguous | one column | 0.03 MB | 2.55 | 12 | 0.042x |
+
+### After: parallel cached decode, fewer copies (full reads)
+
+Full-read times, old and new binaries run alternately at the same moment (this
+machine's absolute speed drifts over a long session, so only same-moment
+comparisons mean anything):
+
+| layout (64 MB `f64`) | before | after |
+|---|---:|---:|
+| chunked + deflate | 110 ms | 69 ms |
+| chunked | 72 ms | 60 ms |
+| contiguous | 56 ms | 30 ms |
+
+What changed: the facade's cached read path decompressed chunks one at a time
+(only the uncached reader was parallel) and pushed every chunk through a 16 MiB
+cache that a 64 MB read simply churns; it now decodes cache misses in parallel
+batches and caches only datasets that fit. Unfiltered chunks are copied
+straight from the file bytes instead of via two intermediate buffers. A
+contiguous dataset is converted straight from the file bytes (one copy instead
+of two), and the native-endian conversions no longer zero a buffer they are
+about to overwrite.
+
+## Search harness baseline (v2.3.0)
+
+Produced by `cargo run --release -p clawhdf5-bench --bin search_harness -- --full`
+on deterministic **clustered** synthetic data (384-dim, unit-normalised; points =
+cluster centre + noise — uniform random vectors are nearly equidistant in high
+dimension and say nothing about embeddings). Recall is measured against an exact
+brute-force scan, 200 queries. This is the *before* picture for the search
+hot-path work; every change to that path should be justified by a re-run.
+
+Two things stand out:
+
+* **HNSW recall does not respond to `ef`** and degrades sharply with size
+  (0.87 → 0.67 → 0.31 recall@10 at 1K / 10K / 100K). Latency plateaus at the same
+  point, i.e. the search exhausts the nodes it can reach: on clustered data the
+  graph is poorly connected. The index selects neighbours by plain top-M
+  distance rather than the HNSW paper's diversity heuristic.
+* **End-to-end `hybrid_search` is ~1000x slower than its vector stage** (49 ms
+  vs ~0.03 ms at 10K; 884 ms at 100K). Each query rebuilds the BM25 index from
+  scratch and rewrites the whole `.h5` file. The first query after `open()`
+  additionally rebuilds the HNSW index (10.5 s at 100K).
+
+### HNSW, N = 1000, dim = 384, M = 16, ef_construction = 64
+
+build: 72.4 ms (13818 vectors/s) · exact scan: 3854 QPS, p50 258 µs
+
+| ef | recall@10 | QPS | p50 µs | p99 µs |
+|---:|---:|---:|---:|---:|
+| 16 | 0.8710 | 59484 | 16 | 31 |
+| 32 | 0.8730 | 46302 | 21 | 25 |
+| 64 | 0.8730 | 31683 | 31 | 44 |
+| 128 | 0.8730 | 24715 | 40 | 49 |
+| 256 | 0.8730 | 24788 | 40 | 50 |
+
+### HNSW, N = 10000, dim = 384, M = 16, ef_construction = 64
+
+build: 802.5 ms (12461 vectors/s) · exact scan: 418 QPS, p50 2363 µs
+
+| ef | recall@10 | QPS | p50 µs | p99 µs |
+|---:|---:|---:|---:|---:|
+| 16 | 0.6695 | 44031 | 19 | 51 |
+| 32 | 0.6705 | 45066 | 22 | 30 |
+| 64 | 0.6705 | 32746 | 30 | 41 |
+| 128 | 0.6705 | 27542 | 36 | 51 |
+| 256 | 0.6705 | 27754 | 36 | 49 |
+
+### HNSW, N = 100000, dim = 384, M = 16, ef_construction = 64
+
+build: 9752.6 ms (10254 vectors/s) · exact scan: 40 QPS, p50 24648 µs
+
+| ef | recall@10 | QPS | p50 µs | p99 µs |
+|---:|---:|---:|---:|---:|
+| 16 | 0.3085 | 18046 | 57 | 84 |
+| 32 | 0.3110 | 21621 | 43 | 75 |
+| 64 | 0.3130 | 20015 | 49 | 70 |
+| 128 | 0.3135 | 15822 | 63 | 99 |
+| 256 | 0.3135 | 15308 | 66 | 124 |
+
+### End to end: `HDF5Memory::hybrid_search` (k = 10, weights 0.7 / 0.3)
+
+| N | ingest ms | checkpoint ms | open ms | first query ms | p50 ms | p99 ms | QPS |
+|---:|---:|---:|---:|---:|---:|---:|---:|
+| 1000 | 11 | 3.9 | 0.9 | 68.1 | 5.48 | 5.57 | 182.5 |
+| 10000 | 114 | 32.2 | 10.9 | 845.0 | 48.56 | 78.65 | 19.8 |
+| 100000 | 1486 | 713.0 | 354.5 | 10486.5 | 883.51 | 975.23 | 1.1 |
+wrote /tmp/claude-1000/-home-osobh-projects-clawhdf5/422f755e-dd25-4c35-8613-5439087e3aaa/scratchpad/baseline_full.json
+
+### After: HNSW neighbour-selection heuristic
+
+Same harness, same data, after replacing closest-M neighbour selection with the
+HNSW paper's diversity heuristic (Algorithm 4, keeping pruned connections) for
+both new links and back-link pruning. Recall@10 at `ef = 64`: **0.87 → 1.00**
+(1K), **0.67 → 1.00** (10K), **0.31 → 0.98** (100K), and it now rises with
+`ef` as it should. The cost is a slower build (extra distance evaluations per
+insert: ~3.5x at 10K); the distance-kernel work that follows targets that.
+
+### HNSW, N = 1000, dim = 384, M = 16, ef_construction = 64
+
+build: 221.3 ms (4519 vectors/s) · exact scan: 3851 QPS, p50 258 µs
+
+| ef | recall@10 | QPS | p50 µs | p99 µs |
+|---:|---:|---:|---:|---:|
+| 16 | 0.9990 | 54760 | 18 | 29 |
+| 32 | 1.0000 | 40422 | 24 | 44 |
+| 64 | 1.0000 | 27744 | 36 | 51 |
+| 128 | 1.0000 | 13164 | 74 | 106 |
+| 256 | 1.0000 | 6879 | 144 | 175 |
+
+### HNSW, N = 10000, dim = 384, M = 16, ef_construction = 64
+
+build: 2733.5 ms (3658 vectors/s) · exact scan: 423 QPS, p50 2362 µs
+
+| ef | recall@10 | QPS | p50 µs | p99 µs |
+|---:|---:|---:|---:|---:|
+| 16 | 0.9975 | 31321 | 27 | 61 |
+| 32 | 1.0000 | 32427 | 29 | 48 |
+| 64 | 1.0000 | 22738 | 42 | 62 |
+| 128 | 1.0000 | 10055 | 99 | 129 |
+| 256 | 1.0000 | 4649 | 214 | 266 |
+
+### HNSW, N = 100000, dim = 384, M = 16, ef_construction = 64
+
+build: 36472.8 ms (2742 vectors/s) · exact scan: 40 QPS, p50 24644 µs
+
+| ef | recall@10 | QPS | p50 µs | p99 µs |
+|---:|---:|---:|---:|---:|
+| 16 | 0.9235 | 11394 | 82 | 194 |
+| 32 | 0.9675 | 12788 | 73 | 161 |
+| 64 | 0.9840 | 10406 | 91 | 186 |
+| 128 | 0.9990 | 7633 | 126 | 248 |
+| 256 | 0.9990 | 2823 | 352 | 510 |
+
+### After: persistent keyword index, no store rewrite per query
+
+`hybrid_search` used to rebuild the BM25 index from scratch (re-tokenising every
+record) and rewrite the whole `.h5` file on **every query**. The index is now
+kept for the life of the store and updated incrementally, and activation boosts
+are persisted by the next checkpoint instead of inside the query. Steady-state
+p50: **5.5 → 0.24 ms** (1K), **49 → 2.1 ms** (10K), **884 → 23 ms** (100K).
+
+The first query after `open()` is slower than before (it pays for the better —
+slower — HNSW build plus the one-off keyword index build); persisting the HNSW
+index removes that.
+
+### End to end: `HDF5Memory::hybrid_search` (k = 10, weights 0.7 / 0.3)
+
+| N | ingest ms | checkpoint ms | open ms | first query ms | p50 ms | p99 ms | QPS |
+|---:|---:|---:|---:|---:|---:|---:|---:|
+| 1000 | 11 | 3.8 | 0.9 | 195.9 | 0.24 | 0.27 | 4130.4 |
+| 10000 | 104 | 31.1 | 10.9 | 2627.1 | 2.09 | 2.11 | 479.5 |
+| 100000 | 1436 | 684.7 | 278.0 | 36308.1 | 22.90 | 25.46 | 43.5 |
+
+### After: vector index persisted with the checkpoint
+
+The HNSW graph (not the vectors, which the store already holds) is saved to
+`<store>.h5.ann` at each checkpoint and reloaded by `open()`, tied to that
+checkpoint by a generation id. The index is now built once per store (the *cold
+index build* column — the first query ever), not once per session. First query
+after `open()`: **196 → 1.7 ms** (1K), **2627 → 15 ms** (10K),
+**36308 → 159 ms** (100K); what remains is the one-off keyword index build.
+Batch saves no longer force a full rebuild either: appended records join the
+index incrementally.
+
+| N | ingest ms | cold index build ms | checkpoint ms | open ms | first query after open ms | p50 ms | p99 ms | QPS |
+|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| 1000 | 14 | 220 | 6.1 | 1.2 | 1.7 | 0.24 | 0.27 | 4049.9 |
+| 10000 | 120 | 2916 | 33.1 | 14.0 | 15.4 | 2.15 | 3.30 | 421.2 |
+| 100000 | 1591 | 40515 | 747.3 | 324.7 | 158.9 | 23.07 | 30.42 | 41.3 |
+
+### After: unit-vector dot product, reusable visited set
+
+Cosine distance recomputed both vector norms on every evaluation; the index now
+stores unit vectors and uses a plain dot product. The per-call `HashSet` of
+visited nodes became a reusable epoch-stamped array. Recall is unchanged.
+Build: **2.75 -> 1.89 s** (10K), **~38 -> 21 s** (100K). QPS at `ef = 64`:
+**22.7K -> 39K** (10K), **10.4K -> 14K** (100K).
+
+### HNSW, N = 1000, dim = 384, M = 16, ef_construction = 64
+
+build: 113.4 ms (8821 vectors/s) · exact scan: 4375 QPS, p50 225 µs
+
+| ef | recall@10 | QPS | p50 µs | p99 µs |
+|---:|---:|---:|---:|---:|
+| 16 | 0.9990 | 144379 | 7 | 15 |
+| 32 | 1.0000 | 110654 | 9 | 17 |
+| 64 | 1.0000 | 80446 | 12 | 25 |
+| 128 | 1.0000 | 38220 | 26 | 36 |
+| 256 | 1.0000 | 20041 | 50 | 62 |
+
+### HNSW, N = 10000, dim = 384, M = 16, ef_construction = 64
+
+build: 1519.4 ms (6581 vectors/s) · exact scan: 422 QPS, p50 2368 µs
+
+| ef | recall@10 | QPS | p50 µs | p99 µs |
+|---:|---:|---:|---:|---:|
+| 16 | 0.9975 | 54608 | 15 | 45 |
+| 32 | 1.0000 | 66009 | 14 | 24 |
+| 64 | 1.0000 | 49854 | 19 | 31 |
+| 128 | 1.0000 | 22403 | 45 | 57 |
+| 256 | 1.0000 | 10096 | 100 | 120 |
+
+### HNSW, N = 100000, dim = 384, M = 16, ef_construction = 64
+
+build: 21084.6 ms (4743 vectors/s) · exact scan: 39 QPS, p50 24739 µs
+
+| ef | recall@10 | QPS | p50 µs | p99 µs |
+|---:|---:|---:|---:|---:|
+| 16 | 0.9235 | 15139 | 61 | 154 |
+| 32 | 0.9675 | 18181 | 53 | 121 |
+| 64 | 0.9840 | 13980 | 70 | 139 |
+| 128 | 0.9990 | 10959 | 86 | 174 |
+| 256 | 0.9990 | 3731 | 254 | 697 |
+
+
+### After: unranked keyword scores, top-k merge (rankings unchanged)
+
+A fusion study (`search_harness --fusion-study`) showed that capping the
+keyword candidate pool is **not** a safe optimisation: against the current
+full-corpus normalisation the final top-10 overlap is only 0.83-0.92 and the
+first result changes for 10-35% of queries, for only a 2x saving. So the fusion
+semantics were left alone and the same answer made cheaper: fusion needs every
+keyword score but not their ranking, so BM25 now returns them unsorted from a
+dense accumulator (it hashed every posting and then sorted every match), and
+the merge selects its top k instead of sorting every candidate. Steady-state
+p50: **0.24 -> 0.07 ms** (1K), **2.1 -> 0.49 ms** (10K), **23 -> 4.65 ms**
+(100K) — **79x / 100x / 190x** faster than the v2.3.0 baseline, with identical
+results.
+
+### End to end: `HDF5Memory::hybrid_search` (k = 10, weights 0.7 / 0.3)
+
+| N | ingest ms | cold index build ms | checkpoint ms | open ms | first query after open ms | p50 ms | p99 ms | QPS |
+|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| 1000 | 11 | 112 | 4.0 | 1.1 | 1.4 | 0.07 | 0.08 | 14077.9 |
+| 10000 | 104 | 1487 | 33.8 | 13.7 | 13.9 | 0.49 | 0.51 | 2020.9 |
+| 100000 | 1376 | 20285 | 728.9 | 353.1 | 142.2 | 4.65 | 4.78 | 214.7 |
+
+### After: batched bulk build (optionally parallel); deletions handled in search
+
+Profiling showed **90% of a build's distance evaluations are in back-link
+pruning**. The bulk build now inserts in batches: plan each node's neighbours
+against the graph as it stood at the start of the batch, link, then prune every
+overflowing list once. That is less work even single-threaded (a node gaining
+several back-links in a batch is pruned once), and with the `parallel` feature
+planning and pruning run on a thread pool. The graph is deterministic and the
+same with or without the feature. Parallelising *within* one insert was tried
+first and gave only 1.45x on 16 cores (tasks too small).
+
+| build | 1K | 10K | 100K |
+|---|---:|---:|---:|
+| v2.4.0 | 116 ms | 1676 ms | ~21 s |
+| batched | 83 ms | 1074 ms | 19.2 s |
+| batched + `parallel` (16 cores) | 34 ms | 388 ms | 5.9 s |
+
+Recall on clustered data is unchanged or slightly better (100K, `ef = 64`:
+0.984 -> 0.9945). On uniform random data it dips slightly (10K, `ef = 64`:
+0.474 -> 0.444), the cost of batch members not seeing each other while
+planning; batches are capped at 1/16 of the graph and 512 nodes.
+
 ## Vector Search Latency

 Brute-force cosine similarity over 384-dimensional embeddings (OpenAI text-embedding-3-small size).
@@ -276,6 +672,102 @@ Session-level:
 | Vector only | 85.4% | 94.2% | 96.6% | 0.8901 |
 | Hybrid | **88.2%** | **95.8%** | **97.8%** | **0.9158** |

+### Fusion method — weighted vs. RRF, full haystack, n=500
+
+Reciprocal rank fusion has been in the codebase since early on but was only
+reachable as a free function over a linear scan, so it had never been compared
+with the weighted sum on equal terms. `HDF5Memory::hybrid_search_with` now
+takes a `Fusion`, and both run over the same HNSW + BM25 candidates:
+
+| Mode | turn Hit@1 | Hit@5 | Hit@10 | MRR | session Hit@1 | session MRR |
+|---|---|---|---|---|---|---|
+| BM25 only | **53.8%** | 75.0% | 81.6% | 0.6320 | 86.2% | 0.8948 |
+| Vector only | 36.0% | 71.8% | 81.6% | 0.5031 | 85.4% | 0.8901 |
+| **Weighted 0.4 / 0.6** | 51.6% | **81.4%** | **87.8%** | **0.6430** | **91.0%** | **0.9347** |
+| RRF (k=60) | 45.0% | 78.8% | 87.6% | 0.5967 | 89.6% | 0.9253 |
+
+**RRF loses to the tuned weighted sum** — 6.6pp of turn Hit@1 and 0.046 of MRR
+— and lands almost exactly where the old `0.7/0.3` weighting did (44.2% / 
+0.5856). That is not a coincidence: RRF combines the two stages by rank with
+*equal* influence, and on this corpus the stages are not equally good. BM25
+alone beats the vector stage by 17.8pp at Hit@1, so any scheme that treats them
+as peers gives up rank-1 accuracy, and RRF discards the score magnitudes that
+would say which stage to believe.
+
+This is a property of the corpus, not a defect in RRF: its selling point is
+robustness when the two stages' scores are not comparable and there is no
+labelled data to tune against. Here there is, so the weighted sum is kept as
+the default. `Fusion::Rrf` remains available for callers whose stages are more
+evenly matched.
+
+### Keyword tokenizer — stemming, full haystack, n=500
+
+The keyword stage lowercases and splits on non-alphanumerics, with no stemming,
+so "training" and "trains" are unrelated terms. `TokenFilter::Stemmed` strips
+common English inflections (plurals, `-ing`/`-ed`, with consonant un-doubling)
+from documents and queries alike. Turn-level:
+
+| Mode | Hit@1 | Hit@5 | Hit@10 | MRR | session Hit@1 |
+|---|---|---|---|---|---|
+| BM25 only | **53.8%** | 75.0% | 81.6% | 0.6320 | 86.2% |
+| BM25 only, stemmed | 52.0% | 77.8% | 84.0% | 0.6320 | 88.0% |
+| Hybrid 0.4/0.6 | 51.6% | **81.4%** | 87.8% | **0.6430** | 91.0% |
+| Hybrid 0.4/0.6, stemmed | 50.2% | **81.4%** | **88.2%** | 0.6394 | **91.4%** |
+
+**Stemming is a trade, not a win, and the default stays off.** It reliably buys
+depth and costs the top rank: on BM25 alone, +2.8pp Hit@5 and +2.4pp Hit@10 for
+−1.8pp Hit@1, with MRR unchanged to four decimal places — the gains deeper down
+exactly offset the loss at rank 1. That is what conflation does: merging
+"train"/"training"/"trains" surfaces documents an exact-match query would never
+reach, and also lets a near-miss outrank the exact hit.
+
+On the configuration that actually ships (hybrid 0.4/0.6) the trade is
+narrower still — Hit@5 identical, Hit@10 +0.4pp, Hit@1 −1.4pp, MRR −0.004 —
+because the vector stage already supplies much of the recall stemming would
+add. There is no case here for changing the default; `TokenFilter::Stemmed`
+is available via `HDF5Memory::set_token_filter` for callers who want Hit@5/@10
+over rank-1 precision.
+
+### Re-ranking and recency — full haystack, n=500
+
+`reranker::rerank` combines temporal decay, source authority and Hebbian
+activation. Until now its combined score contained **no relevance term at
+all** — `RerankInput` did not carry the retrieval score — so a caller that
+re-ranked its candidates threw the retriever's ordering away and returned them
+ordered by age. The OpenClaw backend did exactly that on every search.
+
+Measuring that is unambiguous. "Recency" below is the share of
+`knowledge-update` questions where the newest gold session outranked the stale
+one (see `newest_gold_first`); ~45% is chance.
+
+| Mode | Hit@1 | Hit@5 | Hit@10 | MRR | recency |
+|---|---|---|---|---|---|
+| Hybrid 0.4/0.6, no re-rank | 51.6% | **81.4%** | 87.8% | 0.6430 | 45.0% |
+| + re-rank, **metadata only** (pre-fix) | 11.0% | 24.8% | 43.8% | 0.1829 | **87.5%** |
+| + re-rank, relevance-led, half-life 1 day | **52.0%** | 79.8% | 87.8% | 0.6403 | 51.7% |
+| + re-rank, relevance-led, half-life 7 days | 51.8% | 80.8% | 87.6% | **0.6437** | **52.2%** |
+| + re-rank, relevance-led, half-life 30 days | 51.8% | 81.0% | 87.8% | 0.6427 | 51.4% |
+| + re-rank, relevance-led, half-life 90 days | **52.0%** | 80.4% | 87.8% | 0.6425 | 50.8% |
+
+**The pre-fix row is the finding.** Ordering candidates by recency alone costs
+40.6pp of Hit@1 and two thirds of MRR: the results are the newest memories in
+the pool rather than the ones that answer the question. It does ace the recency
+metric, which is exactly what makes that metric worth having — a number that
+only goes up when a change is good would not have caught this.
+
+With relevance leading, retrieval is preserved (Hit@1 +0.4pp, MRR −0.003
+against no re-ranking) and recency discrimination gains 6–7pp. That is a real
+improvement but not a solved problem: recency only breaks near-ties, so it
+cannot reach the 87.5% the degenerate ordering gets. Those two rows are the
+ends of a trade-off, and the default sits deliberately near the relevance end.
+
+**Half-life is not a sensitive knob.** Across 1, 7, 30 and 90 days recency
+moves 1.4pp and MRR 0.003 — inside the noise of a 500-question run — because
+the temporal term is capped by its weight (0.3) while relevance differences
+between candidates are larger. The 24-hour default is kept; there is no
+measured reason to change it, and a corpus-matched value is not the lever it
+looks like.
+
 ### Weight sweep — full haystack, n=500

 `0.7/0.3` was a documented default, never a searched one. Sweeping
@@ -318,7 +810,8 @@ BM25 at Hit@1. Both dominate `0.7/0.3`.

 The rows below are kept at the three original settings because they are what the
 mode ablation measured — read them as "the shape of each stage in isolation",
-and take the operating point from the sweep.
+and take the operating point from the sweep. `0.4/0.6` is now the shipped
+default (`hybrid::DEFAULT_FUSION`).

 The same pattern shows up independently in omni-cortex's four-signal RRF ablation,
 where adding BM25 to a dense retriever raised nDCG@5 while lowering Hit@1 and MRR.
@@ -1,6 +1,427 @@
 # Changelog

-## Unreleased
+## v2.6.0 (2026-09-20)
+
+### Upgrade Notes
+- **Re-ranked results change, substantially for the better.** `RerankInput`
+  and `ReRankConfig` gained fields (`relevance`, `relevance_weight`), so
+  literal constructions need updating; `..Default::default()` does not. Any
+  caller that re-ranked was previously getting results ordered by age with the
+  retrieval score discarded — see below.
+- **Breaking:** `MemoryCache::embeddings` is a `cache::Embeddings` rather than
+  a `Vec<Vec<f32>>` (indexing still yields a `&[f32]` row); `embeddings_flat`
+  is gone, replaced by `flat_embeddings()`; `rebuild_flat()` is a deprecated
+  no-op.
+- `MemoryConfig` gained `quantized_index` (default `false`, so behaviour is
+  unchanged unless you opt in); literal constructions need the field.
+
+### Retrieval quality
+- `clawhdf5-agent`: **re-ranking discarded the retrieval score.**
+  `reranker::rerank` built its combined score from temporal decay, source
+  authority and Hebbian activation only — `RerankInput` had no relevance field
+  — so re-ranking a candidate pool reordered it by age and threw the
+  retriever's ordering away. The OpenClaw backend re-ranked every search, so
+  this was its shipping behaviour: measured over the full LongMemEval haystack
+  it cost **40.6pp of Hit@1** (11.0% vs 51.6%) and two thirds of MRR (0.183 vs
+  0.643). `RerankInput::relevance` and `ReRankConfig::relevance_weight` (1.0 by
+  default) fix it: relevance leads and the metadata signals break near-ties,
+  which restores retrieval (Hit@1 +0.4pp vs no re-ranking) and improves
+  recency discrimination by 6–7pp. **Breaking:** `RerankInput` and
+  `ReRankConfig` gained fields, so literal constructions need updating;
+  `..Default::default()` does not.
+- `clawhdf5-bench`: the LongMemEval harness feeds the dataset's real session
+  dates to the store instead of a synthetic counter (decay needs true
+  intervals, not just the right order), and reports `newest_gold_first` — on a
+  `knowledge-update` question, did the newest gold session outrank the stale
+  one it supersedes? Plain recall cannot see this, because both are labelled
+  gold. New `--rerank-sweep`.
+
+### Memory
+- `clawhdf5-agent`: **`MemoryConfig::quantized_index`** stores the vector
+  index's own copy of the embeddings as `i8` rather than `f32`, which at 100k
+  384-dim entries takes the index from 266 to 123 MiB and the whole reopened
+  store from 399 to 256 MiB (2.72x -> **1.74x** the raw vectors). Quantised
+  distances are approximate and `ef` cannot compensate — recall@10 tops out at
+  0.967 against f32's 0.9995 — so the query path re-scores the candidate pool
+  against the exact embeddings the store already holds, which restores recall
+  (0.9940 vs 0.9945 at ef=64) for about 13% of QPS. **Off by default**: it
+  trades query speed for memory, and which side is worth more depends on the
+  deployment. The setting is persisted, so a reopened store does not silently
+  revert to four times the index memory.
+- `clawhdf5-ann`: `Storage::Int8` and the `build_with` / `new_with` /
+  `from_graph_bytes_with` constructors that select it. The scale is per row,
+  not global — a fixed `[-1, 1]` scale spends fewer than 12 of the 255 levels
+  on a unit-length 128-dim vector and is unusable (0.35 top-10 overlap against
+  an exact ranking, versus 0.99 per row). `compact()` keeps the storage it was
+  given; serialized indexes still carry f32 vectors, so a quantised index is
+  rebuilt rather than loaded.
+- `clawhdf5-agent`: **a loaded store holds ~30% less memory** (100k 384-dim
+  entries: 505 -> 357 MiB, 3.44x -> 2.43x the raw vectors). The cache kept
+  every embedding twice — a `Vec<Vec<f32>>` and a flattened copy for the
+  batched kernels, maintained in lock-step — so it now stores only the flat
+  buffer and indexes into it. Recall and query latency are unchanged.
+  **Breaking:** `MemoryCache::embeddings` is a `cache::Embeddings` rather than
+  a `Vec<Vec<f32>>` (indexing still yields a `&[f32]` row); `embeddings_flat`
+  is gone, replaced by `flat_embeddings()`; `rebuild_flat()` is a deprecated
+  no-op. Rows are now always exactly `dim` long — shorter ones are
+  zero-padded — which makes the ragged-row case that used to silently
+  misalign the flattened copy unrepresentable.
+- `clawhdf5-bench`: `search_harness --footprint` reports live heap use per
+  stage, measured with a counting allocator (RSS cannot see a structure freed
+  into the allocator's own pool).
+
+### Testing
+- The Python interop suites honour **`CLAWHDF5_PYTHON`**, and `ci-test.sh`
+  picks up a `.venv/bin/python` automatically. On a PEP 668 "externally
+  managed" system h5py cannot be installed into the system interpreter at all,
+  so every interop suite — the h5py writer round-trips, the facade, netCDF4
+  and the reference files — was skipping silently. A silent skip here is
+  exactly how the v5 compound-datatype bug reached a release.
+  `CLAWHDF5_REQUIRE_INTEROP=1` still turns a skip into a failure.
+
+## v2.5.0 (2026-09-19)
+
+### Upgrade Notes
+- **Retrieval rankings change, for the better.** The default fusion weights
+  move from `0.7/0.3` to `0.4/0.6` (`hybrid::DEFAULT_FUSION`), measured over the
+  full LongMemEval haystack: turn-level Hit@1 51.6% vs 44.2%, MRR 0.643 vs
+  0.586. `unified_search` and the OpenClaw backend pick this up automatically;
+  callers passing weights to `hybrid_search` explicitly are unaffected.
+- **Out-of-range selections are now errors.** `read_*_selection` used to return
+  data for a selection that ran past a dataset edge — a hyperslab came back
+  zero-padded, and a point with an out-of-range coordinate wrapped into the
+  next row. Both are now `FormatError::SelectionOutOfBounds`. Code relying on
+  the old (wrong) values will start seeing errors.
+- **Large compressed datasets written without explicit chunk dimensions get a
+  different layout.** They used to be stored as one chunk; they are now split
+  to ~1 MiB chunks. The files stay standard and h5py-readable, and explicit
+  `with_chunks` is unaffected.
+- `rayon` is now a default dependency of `clawhdf5-agent` (the parallel index
+  build). Opt out with `--no-default-features --features float16,hnsw`.
+- `clawhdf5-ann` search results no longer shrink when records near the query
+  have been deleted, so a search that previously returned fewer than `k`
+  results now returns `k`.
+
+### Retrieval quality
+- `clawhdf5-agent`: optional keyword stemming — `bm25::TokenFilter::Stemmed`
+  and `HDF5Memory::set_token_filter`, so "training" and "trains" match. **Off
+  by default**, on measurement rather than principle: over the full LongMemEval
+  haystack it buys depth and costs the top rank (BM25 alone: Hit@5 +2.8pp,
+  Hit@10 +2.4pp, Hit@1 −1.8pp, MRR unchanged), and on the shipping hybrid
+  configuration the trade is narrower still. See `BENCHMARKS.md`.
+- `clawhdf5-agent`: **`QueryExpander::expand` panicked on ordinary non-ASCII
+  input** — `"İ AI"` was enough. It searched a lowercased copy of the query and
+  then sliced the *original* with those offsets, which only works while
+  lowercasing preserves byte length (Turkish `İ` is 2 bytes and lowercases to
+  3). Depending on where the offsets drifted it either corrupted the output
+  ("İstanbul AI trip" lost a character) or panicked. Matching now walks the
+  original string.
+- `clawhdf5-agent`: query expansion no longer rewrites text inside words.
+  `replace_word_case_insensitive` did a plain substring replace despite its
+  name, so "training" became "trArtificial Intelligencening" and "programming"
+  became "Pull Requestogramming" — every acronym expansion of ordinary prose
+  was corrupt. Matches now require word boundaries; genuine acronyms
+  (`API`, `database`) still expand.
+- `clawhdf5-agent`: **the default fusion weights are now the measured ones.**
+  A sweep of every 0.1 step over the full LongMemEval haystack (500 questions,
+  real MiniLM embeddings) shows the long-standing `0.7/0.3` default is
+  *strictly dominated* by `0.4/0.6` — turn-level Hit@1 51.6% vs 44.2%, Hit@5
+  81.4% vs 79.2%, Hit@10 87.8% vs 85.8%, MRR 0.643 vs 0.586, and better at
+  session level too. The finding was recorded in `BENCHMARKS.md` but had never
+  been applied: `unified_search` and the OpenClaw backend both hardcoded
+  `0.7/0.3`. They now use `hybrid::DEFAULT_FUSION`. **Callers passing weights
+  to `hybrid_search` explicitly are unaffected** — pass `0.4`/`0.6` (or use
+  `hybrid_search_with`) to get the tuned behaviour.
+- `clawhdf5-agent`: fusion is now selectable. New `hybrid::Fusion`
+  (`Weighted { vector, keyword }` or `Rrf { k }`), `hybrid::fuse`,
+  `hybrid::hybrid_search_fused` and `HDF5Memory::hybrid_search_with`.
+  Reciprocal rank fusion existed but was unreachable from the store, so it had
+  never been measured against the weighted sum; the LongMemEval bench now has
+  an `RRF` mode.
+
+### HDF5 Read Path
+- **Selection reads cost what the selection costs.** `read_*_selection` decoded
+  the *entire* dataset and then picked elements out, so a 64 x 64 window of a
+  64 MB compressed dataset took 105 ms - about as long as reading all of it.
+  Now only the rows (contiguous) or chunks that overlap the selection's
+  bounding box are read and decompressed: that window takes 0.39 ms, one row
+  2.7 ms, one column 5.2 ms. Results are identical to the full-read path
+  (equivalence-tested over random hyperslabs and point lists, ranks 1-3,
+  contiguous / chunked / deflate). New `read_harness` bench binary.
+- **Faster full reads** (same-moment A/B, 64 MB `f64`): chunked + deflate
+  110 -> 69 ms, chunked 72 -> 60 ms, contiguous 56 -> 30 ms. The facade's
+  cached read path now decompresses cache misses in parallel batches (it was
+  sequential; only the uncached reader was parallel) and caches only datasets
+  that fit the chunk cache; unfiltered chunks are copied straight from the file
+  bytes; a contiguous dataset is converted straight from the file bytes; and
+  the native-endian conversions no longer zero a buffer before overwriting it.
+- **Datasets indexed by a version-2 B-tree now read** (layout v4, chunk index
+  type 5 — what `libver='latest'` uses for two or more unlimited dimensions;
+  previously "unsupported chunked layout"). The four copies of the chunk-index
+  dispatch are now one shared function, so every read path gets it.
+- **`H5T_STD_REF` references** (HDF5 1.12+, datatype message version 4) parse:
+  `ReferenceType` gains `Object2`, `DatasetRegion2` and `Attribute`, and
+  `read_object_references` decodes the new object references. Previously any
+  dataset of this type failed with `InvalidReferenceType(2)`. Tested against a
+  file written by HDF5 2.0 itself (fixture + generator script committed).
+- **Automatic chunk sizes.** Asking for compression (or any filter) without
+  `with_chunks` used to store the whole dataset as one chunk, so any read had
+  to decompress everything and nothing could be decoded in parallel. Datasets up
+  to 1 MiB stay a single chunk, as before; larger ones are split by halving the
+  dimensions in turn until a chunk is at most 1 MiB (the approach h5py takes).
+  **Behaviour change:** large compressed datasets written without explicit
+  chunk dimensions get a different (standard, h5py-readable) layout. Explicit
+  `with_chunks` is unaffected.
+- **Out-of-range selections are errors.** They used to return data: a hyperslab
+  past an edge came back padded with zeros, and a point whose column was out of
+  range wrapped into the next row and returned that element. Now
+  `FormatError::SelectionOutOfBounds` (also for a rank mismatch or overlapping
+  blocks).
+
+### Search
+- `clawhdf5-ann`: **faster index builds.** Back-link pruning is 90% of a
+  build's distance evaluations; the bulk build now inserts in batches and
+  prunes each overflowing neighbour list once per batch (10K: 1676 -> 1074 ms).
+  With the `parallel` feature, planning and pruning run on a thread pool (10K:
+  388 ms, 100K: ~21 s -> 5.9 s on 16 cores). The graph is deterministic and
+  identical with or without the feature. `clawhdf5-agent`'s `parallel` feature
+  enables it for the agent's index and is now **on by default** (adds `rayon`
+  to the default dependency set; build with `--no-default-features --features
+  float16,hnsw` to opt out).
+- `clawhdf5-ann`: `HnswIndex::search` returned fewer than `k` results — often
+  none — when the records nearest the query had been deleted: it collected `ef`
+  candidates, *then* dropped the deleted ones, *then* took `k`. Deleted nodes
+  are now traversed as waypoints but never occupy a result slot, so a search
+  returns the `k` nearest live records. Matters for any store that deletes or
+  supersedes memories without compacting straight away.
+
+## v2.4.0 (2026-09-19)
+
+### Upgrade Notes
+- **Search results improve on upgrade.** The HNSW index now reaches true
+  neighbours it previously could not (recall@10 0.31 -> 0.98 at 100K records on
+  clustered data), so `hybrid_search` rankings change for the better. The agent
+  rebuilds its index from the store automatically; a standalone `HnswIndex`
+  persisted with `to_hdf5_bytes` keeps its old graph until rebuilt.
+- **`hybrid_search` no longer writes the store.** Hebbian activation boosts are
+  persisted by the next checkpoint (any flushing write, `flush_wal`, or when
+  the `HDF5Memory` is dropped) instead of inside every query; a crash before
+  then forgets only the boosts since the last checkpoint. Activation weights
+  are now capped at 16.
+- A new sidecar file, `<store>.h5.ann`, holds the vector index graph. It is
+  derived data: safe to delete (the index is rebuilt), copied by `snapshot()`,
+  and worth including when copying a store by hand to avoid a rebuild.
+- `BM25Index` no longer caches IDF and gained `add_document`,
+  `remove_document`, `pad_to`, `scores`, `len` and `is_empty`; results are now
+  deterministic (ties break by record id).
+
+### Search
+- `clawhdf5-ann`: **HNSW recall fix.** Neighbours were chosen as the plain
+  closest-M, which on clustered data (what embeddings look like) turns each
+  cluster into an island: recall@10 was 0.87 / 0.67 / 0.31 at 1K / 10K / 100K
+  vectors and did not improve with `ef`. The index now uses the HNSW paper's
+  diversity heuristic (Algorithm 4 with kept pruned connections) when linking a
+  new node and when pruning back-links: recall@10 at `ef = 64` is 1.00 / 1.00 /
+  0.98 and responds to `ef`. Builds are slower (~3.5x at 10K). Existing
+  persisted indexes keep their old graph until rebuilt; the agent rebuilds its
+  index from the cache, so stores pick this up automatically.
+- `clawhdf5-agent`: **`hybrid_search` is 23-39x faster in steady state** (p50
+  5.5 -> 0.24 ms at 1K records, 49 -> 2.1 ms at 10K, 884 -> 23 ms at 100K).
+  Every query used to rebuild the BM25 index from scratch and rewrite the whole
+  `.h5` file. The keyword index now lives for the life of the store and is
+  updated incrementally (add / remove / in-place update, exactly equivalent to
+  a fresh build - property-tested), and a query no longer writes the store.
+  **Behaviour change:** Hebbian activation boosts are persisted by the next
+  checkpoint (any flushing write, `flush_wal`, or drop) rather than
+  immediately; a crash in between forgets only the boosts since the last
+  checkpoint. Activation weights are now capped (16.0) - they grew without
+  bound.
+- `clawhdf5-agent`: **the vector index is persisted**, so `open()` no longer
+  rebuilds it on the first search (first query after open: 2627 -> 15 ms at 10K
+  records, 36 s -> 159 ms at 100K). The HNSW graph — not the vectors, which the
+  store already holds — is written to `<store>.h5.ann` at each checkpoint and
+  tied to it by a generation id in `/meta`; a missing, stale, damaged or
+  structurally invalid sidecar is ignored and the index rebuilt. Records
+  replayed from the WAL join the loaded index incrementally; a replayed update
+  or delete invalidates it. `snapshot()` copies it. Batch saves no longer force
+  a full index rebuild.
+- `clawhdf5-ann`: faster HNSW build and search with identical recall. The
+  cosine metric stores unit vectors and compares them with a plain dot product
+  (it re-derived both norms on every distance evaluation), and the per-call
+  `HashSet` of visited nodes is a reusable epoch-stamped array. Build 2.75 ->
+  1.89 s at 10K and ~38 -> 21 s at 100K; QPS at `ef = 64` 22.7K -> 39K at 10K.
+  Distances returned by `search` are unchanged (1 - cosine). Indexes loaded
+  from older HDF5 files are normalised on load.
+- `clawhdf5-accel`: the SIMD backend is detected once per process instead of
+  on every kernel call.
+- `clawhdf5-ann`: `HnswIndex::graph_to_bytes` / `from_graph_bytes` — graph-only
+  serialization (checksummed, every neighbour id and level validated on load).
+- `clawhdf5-agent`: a further 4-5x on `hybrid_search` with **identical
+  rankings** (p50 now 0.07 / 0.49 / 4.65 ms at 1K / 10K / 100K — 79x / 100x /
+  190x faster than v2.3.0). Fusion needs every keyword score but not their
+  ranking: new `BM25Index::scores` returns them unsorted from a dense
+  accumulator (it hashed every posting, then sorted every match), and
+  `merge_vector_keyword` selects its top k instead of sorting every candidate.
+  Capping the keyword candidate pool was measured and rejected: it changes the
+  top-10 for most queries (`search_harness --fusion-study`).
+- `clawhdf5-agent`: BM25 results are deterministic (ties break by record id),
+  top-k uses a bounded heap, and the "WAND early termination" that computed a
+  bound and then ignored it is gone. IDF is computed per query.
+- `clawhdf5-bench`: new `search_harness` binary — HNSW recall@10 / QPS / latency
+  per `ef` against an exact scan, and end-to-end `hybrid_search` timings, on
+  deterministic clustered (or `--uniform`) data. Baseline in `BENCHMARKS.md`.
+
+## v2.3.0 (2026-09-19)
+
+### Upgrade Notes
+- **A memory store now has a single writer.** `HDF5Memory::create`/`open` take
+  an exclusive lock (`<store>.h5.lock`); a second open of the same store — in
+  the same or another process — returns `MemoryError::Locked`. Code that opened
+  a second handle just to read should use `HDF5Memory::open_read_only`.
+- **Unsigned array attributes arrive as `AttrValue::U64Array`**, not
+  `I64Array`, and `attrs()` may now return `AttrValue::Raw`. Exhaustive matches
+  on `AttrValue` need the two new arms.
+- **WAL header version 3 → 4.** v3 files are read and upgraded in place, but a
+  store written by 2.3.0 with a pending WAL cannot be opened by 2.2.0 or
+  earlier (it is refused, not corrupted). Checkpoint first
+  (`flush_wal`) if you need to downgrade.
+- `MemoryConfig::compression` now uses deflate unless the agent's new `zstd`
+  feature is enabled; it previously failed outright in a default build.
+- `MemoryError` gained `Locked`; `FormatError` gained `UnresolvedSharedMessage`,
+  `ExternalDataFilesUnsupported` and `ExternalLinkUnsupported`; `MessageType`
+  gained `ExternalDataFiles`.
+
+### Bug Fixes
+- `clawhdf5-format`: compound datatypes written with **default libver bounds**
+  (datatype message version 1 — what plain `h5py.File(path, 'w')` produces)
+  were mis-parsed. The v1 member layout carries 28 bytes of legacy array
+  fields after the byte offset (the parser skipped 24), and v2 pads member
+  names to 8 bytes and has no array fields at all (the parser did neither), so
+  every member after the first byte offset was read from the wrong position —
+  typically surfacing as `Overflow("compound member ...")` on read. Found by
+  adding a default-libver axis to the h5py interop tests; byte-level regression
+  tests for v1 and v2 added.
+- `clawhdf5-gpu`: `gpu_tests` could hang forever under the default parallel
+  test runner — every test created its own wgpu instance and device at once.
+  Tests now serialise GPU access, and GPU→CPU readback waits are bounded
+  (30 s) so a wedged driver returns `GpuError::BufferMap` instead of blocking.
+- `clawhdf5-agent`: `benches/bench.rs` and `benches/memory_bench.rs` no longer
+  compiled against the current `strategy`/`consolidation` APIs.
+
+### HDF5 Compatibility
+- `clawhdf5-format`/`clawhdf5`: datasets and attributes that use a **committed
+  (named) datatype** now read correctly. They store a shared-message reference;
+  the facade parsed the reference bytes as the datatype (`Time { size: 0 }`,
+  unreadable data) and silently dropped such attributes. The shared-reference
+  parser itself was wrong for real files: version 2 has no reserved bytes, and
+  the version 3 types were inverted (1 = SOHM heap, 2 = committed).
+- **Fill values are applied on read.** There was no Fill Value message parser:
+  the holes of a sparse chunked dataset read as zeros even when the fill value
+  was not zero (silently wrong data), and a dataset that was created but never
+  written failed with `NoDataAllocated` where h5py returns a filled array.
+  Messages v1–v3 and the old 0x0004 form are parsed; the fill value is written
+  into exactly the chunk-grid cells missing from the chunk index.
+- **Soft links are followed** during path resolution, in old- and new-style
+  groups (absolute/relative targets, links to groups, links through links),
+  with a depth limit so a link cycle is an error rather than a hang. A dangling
+  link reports the target it could not find.
+- Things the reader does not follow are now explicit errors instead of wrong
+  answers: an external link is `ExternalLinkUnsupported { filename,
+  object_path }` (was `PathNotFound`), and a dataset whose raw data lives in
+  external files (message 0x0007, now a known `MessageType`) is
+  `ExternalDataFilesUnsupported` (it would otherwise read as fill values).
+- **`attrs()` no longer drops attributes.** Any attribute whose datatype had
+  no `AttrValue` variant was omitted with no error — including every Python
+  `bool` (h5py stores `attrs["flag"] = True` as an enum), complex numbers,
+  compound values and object references. Now:
+  - numpy/h5py-style booleans (an enum of exactly `FALSE`=0 / `TRUE`=1) decode
+    as `I64` / `I64Array` of 0/1;
+  - new `AttrValue::U64Array` keeps unsigned arrays unsigned (they were cast to
+    `I64Array`, so values above `i64::MAX` came back negative). **Behaviour
+    change:** code matching `I64Array` for an unsigned attribute must also
+    match `U64Array` (the netCDF-4 CF helpers and Python bindings do);
+  - new `AttrValue::Raw { datatype, shape, data }` carries everything else
+    verbatim, decodable with `clawhdf5_format::data_read` against `datatype`.
+  Both new variants are writable, so an attribute can be copied between files
+  unchanged. Python receives `Raw` as `{"dtype", "shape", "data"}`.
+- All of the above are covered by h5py interop tests under both default and
+  `libver='latest'` bounds, compared against h5py's own readback.
+
+### Security
+- `clawhdf5`: virtual-dataset source file names are untrusted input but were
+  joined straight onto the opened file's directory, so a crafted file could
+  make the reader open any path the process can reach (absolute path, or `..`
+  components). Only plain relative paths inside that directory are accepted.
+
+### Durability & Integrity
+- `clawhdf5-agent`: a crash between writing a checkpoint and truncating the WAL
+  no longer **duplicates every pending entry** on the next open. Each
+  checkpoint records a `WalMark` (byte length + chained CRC of the WAL prefix it
+  folded in) in `/meta`; `open()` skips exactly that prefix when it is still
+  present. No WAL format change for this; older files behave as before.
+- `clawhdf5-agent`: checkpoints and snapshots are durable as a unit — the temp
+  file is synced before the rename and the directory after it. Individual WAL
+  appends remain unsynced by design (documented in `CLAUDE.md`).
+- `clawhdf5-agent`: `save_or_update` hits are logged as a new `Update` WAL
+  record, so replay updates in place instead of appending a duplicate. WAL
+  header version 3 → 4 (so older builds refuse the file rather than truncating
+  a record they can't parse); v3 files are read and upgraded in place.
+- `clawhdf5-agent`: loading validates every per-record dataset length (a
+  truncated store is now `MemoryError::Schema`, not a later panic), fixes the
+  `n.len() == n.len()` tautology that trusted a norms dataset of any length,
+  and rejects `embedding_dim == 0` with records present.
+- `clawhdf5-agent`: eight behavioural `MemoryConfig` fields are now persisted in
+  `/meta`. Previously they reset to defaults on every open — a compressed store
+  was rewritten uncompressed, `wal_enabled = false` flipped back to `true`.
+- `clawhdf5-agent`: `compression = true` never worked in a default build (it
+  requested Zstd without enabling the feature, so every checkpoint failed with
+  `unsupported filter: 32015`). Default builds now use deflate; Zstd is the new
+  opt-in `zstd` feature.
+- `clawhdf5-agent`: **single-writer lock** (`<store>.h5.lock`,
+  `MemoryError::Locked`) — two handles on one store used to silently destroy
+  each other's data. New `HDF5Memory::open_read_only` gives a lock-free,
+  never-writing view; the CLI's read-only subcommands use it.
+- `clawhdf5-agent`: an unreadable WAL (torn header / bad magic) is quarantined
+  (`HDF5Memory::quarantined_wal()`) instead of blocking `open()` of a healthy
+  store. A WAL from an unknown newer version still fails and is left intact.
+- `clawhdf5-agent`: provenance records are renumbered on compaction (they
+  weren't, so every later `save_or_update` raised a false High integrity
+  alert); pending anomaly alerts and tracked sessions are bounded;
+  `snapshot()` includes entries still in the WAL.
+- `clawhdf5-agent`: hybrid ranking is deterministic (index tie-breaks instead
+  of `HashMap` order); a set of identical positive scores — including a single
+  candidate — normalises to 1.0 rather than 0.0; the Hebbian boost no longer
+  reinforces zero-score filler results.
+- `clawhdf5-format`: chunked/VDS/hyperslab reads size their buffers with
+  overflow-checked arithmetic and fallible allocation, so crafted dimensions
+  are `FormatError::Overflow` instead of a wrapped size or a process abort;
+  `parallel_read` bounds checks use `checked_add`.
+- `clawhdf5`: a malformed filter-pipeline message is an error instead of being
+  treated as "no filters" (which returned compressed bytes as data);
+  `FileBuilder::write` is atomic and synced instead of truncating the
+  destination first.
+
+### CI / Testing
+- CI now lints every target (`cargo clippy --all-targets`) plus
+  `clawhdf5-format`'s optional features, compiles all benches, and tests the
+  format feature matrix. Previously test/bench code and feature-gated modules
+  were never linted; the accumulated clippy backlog is fixed.
+- CI installs python3 + h5py/numpy/netCDF4/xarray and sets
+  `CLAWHDF5_REQUIRE_INTEROP=1`, which turns a missing interop dependency into a
+  test **failure**. Until now every h5py/netCDF4 interop test silently skipped
+  in CI, which is how the HDF5 2.0 compound bug fixed in v2.2.0 reached a user.
+  The `#[ignore]`d `writer_h5py_tests` suite is run explicitly.
+- h5py-generated-file tests now cover default libver bounds as well as
+  `libver='latest'` (HDF5 2.0 raised the default low bound to 1.8).
+- `clawhdf5-agent`: WAL property tests (round trip; after any corruption the
+  entries read back are an exact prefix of what was written — 1500 seeded
+  cases), a crash-recovery matrix (an on-disk image after every operation, the
+  checkpoint window, and the WAL torn at every byte length, each reopened and
+  checked against a model), and a WAL fuzz target.
+- Optional fuzz smoke run (`CLAWHDF5_FUZZ_SECONDS=N scripts/ci-test.sh`); new
+  datatype corpus seeds for v1 compound and native complex messages.
+
+## v2.2.0 (2026-09-18)

 ### Security
 - `clawhdf5-format`: bounded decompression output (`MAX_DECOMPRESS_SIZE`) for
@@ -245,6 +666,16 @@
  reading compound types and — critically — every chunked/compressed dataset
  written by HDF5 2.0. Found by running the h5py interop tests against
  h5py 3.16 / HDF5 2.0.
+  Independently reported (with a patch) against the v2.1.0 tag by
+  M. Scot Breitenfeld (The HDF Group) — v2.1.0 predates this fix.
+- `clawhdf5-format`: parse HDF5 2.0 native complex datatypes (class 11,
+  datatype version 5, e.g. `H5T_COMPLEX_IEEE_F64LE`). The properties are a
+  single base floating-point datatype, not a compound-style member list; the
+  old parser read the base type's bytes as member names, producing a garbage
+  datatype, and failed with `UnexpectedEof` when a complex type was nested in
+  a compound. It is now surfaced as the equivalent `{r, i}` compound (the
+  shape h5py writes for numpy complex dtypes), with a size check against the
+  base type. Validated end-to-end against an HDF5 2.0-written file.

 ### Performance
 - `clawhdf5-format`: chunked writes now compress all chunks up front via
@@ -33,7 +33,66 @@ Cargo workspace with 16 crates under `crates/` (plus `libaec-sys`, an internal F
  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.
- WAL (write-ahead log) for crash-safe persistence, with a CRC32 trailer per entry so a corrupted entry stops replay cleanly instead of loading bad data
+  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). `MemoryConfig::quantized_index` (off by
+  default, persisted) stores the index's own copy of the embeddings as `i8`,
+  which roughly halves a loaded store's memory (2.72x -> 1.74x the raw vectors
+  at 100K); because quantised distances are approximate and `ef` cannot
+  compensate, the query path then re-scores the candidate pool against the
+  exact embeddings, which holds recall at the f32 index's level and costs
+  ~13% of QPS. `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
@@ -21,10 +21,10 @@ members = [
 resolver = "2"

 [workspace.package]
-version = "2.1.0"
+version = "2.6.0"
 edition = "2024"
 license = "MIT"
-repository = "https://github.com/redclawsystems/clawhdf5"
+repository = "https://git.redclaw.dev/quantumclaw/clawhdf5"

 [workspace.dependencies]
 tempfile = "3"
@@ -432,6 +432,13 @@ ClawhDF5's agent memory design draws from 15+ recent papers:
 | `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 |
+
+`MemoryConfig::quantized_index` (off by default) stores the HNSW index's own
+copy of the embeddings as `i8`, roughly halving a loaded store's memory
+(2.72x -> 1.74x the raw vectors at 100k x 384). Quantised distances are
+approximate, so the query path re-scores the candidate pool against the exact
+embeddings the store already holds — recall matches the `f32` index, at about
+13% fewer queries per second. See `BENCHMARKS.md`, "Quantising the index copy".
 | `parallel` | no | Rayon parallel search |
 | `fast-math` | no | BLAS matrix-vector multiply |
 | `accelerate` | no | Apple Accelerate / AMX (macOS) |
@@ -492,6 +499,14 @@ 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
+scripts/ci-test.sh                # what CI runs: fmt, clippy matrix, tests,
+                                  # h5py/netCDF4 interop, no_std
+
+# The interop suites need a Python with h5py; on a PEP 668 system that has to
+# be a virtualenv. `ci-test.sh` finds `.venv` on its own, or set
+# CLAWHDF5_PYTHON. Without one they skip — set CLAWHDF5_REQUIRE_INTEROP=1 to
+# make that a failure instead.
+python3 -m venv .venv && .venv/bin/pip install h5py numpy netCDF4 xarray

 # Benchmarks
 cargo bench -p clawhdf5-agent               # agent memory suite
@@ -1,10 +1,10 @@
 [package]
 name = "clawhdf5-accel"
-version = "2.1.0"
+version = "2.6.0"
 edition = "2024"
 description = "SIMD-accelerated operations for rustyhdf5"
 license = "MIT"
-repository = "https://github.com/redclawsystems/clawhdf5"
+repository = "https://git.redclaw.dev/quantumclaw/clawhdf5"
 readme = "README.md"
 keywords = ["hdf5", "simd", "acceleration", "performance"]
 categories = ["science", "algorithms"]
@@ -111,7 +111,11 @@ pub unsafe fn cosine_similarity(a: &[f32], b: &[f32]) -> f32 {
        }

        let denom = (norm_a * norm_b).sqrt();
-        if denom == 0.0 { 0.0 } else { dot / denom }
+        if denom < f32::EPSILON {
+            0.0
+        } else {
+            dot / denom
+        }
    }
 }

@@ -89,7 +89,11 @@ pub unsafe fn cosine_similarity(a: &[f32], b: &[f32]) -> f32 {
        }

        let denom = (norm_a * norm_b).sqrt();
-        if denom == 0.0 { 0.0 } else { dot / denom }
+        if denom < f32::EPSILON {
+            0.0
+        } else {
+            dot / denom
+        }
    }
 }

@@ -61,8 +61,14 @@ pub enum Backend {
    Scalar,
 }

-/// Detect the best available SIMD backend at runtime.
+/// The best available SIMD backend, detected once per process. Every kernel
+/// dispatches through this, so it sits in the innermost loop of every search.
 pub fn detect_backend() -> Backend {
+    static BACKEND: std::sync::OnceLock<Backend> = std::sync::OnceLock::new();
+    *BACKEND.get_or_init(detect_backend_uncached)
+}
+
+fn detect_backend_uncached() -> Backend {
    #[cfg(target_arch = "aarch64")]
    {
        return Backend::Neon; // Always available on aarch64
@@ -361,6 +367,18 @@ mod tests {
        assert!(approx_eq(cosine_similarity(&a, &b), 0.0, EPSILON));
    }

+    #[test]
+    fn test_cosine_near_zero_norm_clamped() {
+        // denom = 1e-4 * 1e-4 = 1e-8, comfortably below f32::EPSILON
+        // (~1.19e-7) but not exactly 0.0 — must still clamp to 0.0 so
+        // callers computing `1.0 - cosine_similarity(...)` treat these
+        // as maximally dissimilar, matching the pre-SIMD scalar guard.
+        let a = [1e-4f32];
+        let b = [1e-4f32];
+        assert_eq!(cosine_similarity(&a, &b), 0.0);
+        assert_eq!(scalar::cosine_similarity(&a, &b), 0.0);
+    }
+
    #[test]
    fn test_cosine_scalar_vs_dispatch() {
        let a: Vec<f32> = (0..384).map(|i| (i as f32).sin()).collect();
@@ -94,7 +94,11 @@ pub unsafe fn cosine_similarity(a: &[f32], b: &[f32]) -> f32 {
    }

    let denom = (norm_a * norm_b).sqrt();
-    if denom == 0.0 { 0.0 } else { dot / denom }
+    if denom < f32::EPSILON {
+        0.0
+    } else {
+        dot / denom
+    }
 }

 /// NEON L2 distance.
@@ -21,7 +21,11 @@ pub fn cosine_similarity(a: &[f32], b: &[f32]) -> f32 {
        norm_b += y * y;
    }
    let denom = (norm_a * norm_b).sqrt();
-    if denom == 0.0 { 0.0 } else { dot / denom }
+    if denom < f32::EPSILON {
+        0.0
+    } else {
+        dot / denom
+    }
 }

 pub fn batch_cosine(query: &[f32], vectors: &[&[f32]], results: &mut [(usize, f32)]) {
@@ -1,21 +1,21 @@
 [package]
 name = "clawhdf5-agent"
-version = "2.1.0"
+version = "2.6.0"
 edition = "2024"
 description = "HDF5-backed persistent memory store for on-device AI agents"
 license = "MIT"
-repository = "https://github.com/redclawsystems/clawhdf5"
+repository = "https://git.redclaw.dev/quantumclaw/clawhdf5"
 readme = "README.md"
 keywords = ["agent", "memory", "hdf5", "vector-search", "embedding"]
 categories = ["database", "science", "algorithms"]

 [dependencies]
-clawhdf5-format = { path = "../clawhdf5-format", version = "2.1.0", features = ["parallel", "fast-checksum"] }
-clawhdf5 = { path = "../clawhdf5", version = "2.1.0" }
-clawhdf5-io = { path = "../clawhdf5-io", version = "2.1.0", features = ["mmap"] }
-clawhdf5-accel = { path = "../clawhdf5-accel", version = "2.1.0" }
-clawhdf5-ann = { path = "../clawhdf5-ann", version = "2.1.0", optional = true }
-clawhdf5-gpu = { path = "../clawhdf5-gpu", version = "2.1.0", optional = true, default-features = false }
+clawhdf5-format = { path = "../clawhdf5-format", version = "2.6.0", features = ["parallel", "fast-checksum"] }
+clawhdf5 = { path = "../clawhdf5", version = "2.6.0" }
+clawhdf5-io = { path = "../clawhdf5-io", version = "2.6.0", features = ["mmap"] }
+clawhdf5-accel = { path = "../clawhdf5-accel", version = "2.6.0" }
+clawhdf5-ann = { path = "../clawhdf5-ann", version = "2.6.0", optional = true }
+clawhdf5-gpu = { path = "../clawhdf5-gpu", version = "2.6.0", optional = true, default-features = false }
 serde = { workspace = true }
 byteorder = "1"
 half = { workspace = true, optional = true }
@@ -45,9 +45,14 @@ name = "memory_bench"
 harness = false

 [features]
-default = ["float16", "hnsw"]
+default = ["float16", "hnsw", "parallel"]
 float16 = ["half"]
-parallel = ["rayon"]
+# Rayon-parallel brute-force search strategies, and a parallel bulk build of
+# the HNSW index (same graph, several times faster on a multi-core machine).
+parallel = ["rayon", "clawhdf5-ann?/parallel"]
+# Compress embeddings with Zstd instead of deflate when
+# `MemoryConfig::compression` is on. Off by default: it links libzstd (C).
+zstd = ["clawhdf5/zstd"]
 # HNSW approximate-nearest-neighbour acceleration for the vector stage of
 # hybrid_search. On by default; the index is rebuilt from the cache on demand
 # and stays self-consistent with the persisted memory store. Disable with
@@ -483,7 +483,7 @@ fn rayon_benches(c: &mut Criterion) {
                use rayon::prelude::*;
                let query_norm = vector_search::compute_norm(&query);
                let num_cores = rayon::current_num_threads().max(1);
-                let chunk_size = (n + num_cores - 1) / num_cores;
+                let chunk_size = n.div_ceil(num_cores);
                let mut results: Vec<(usize, f32)> = vectors
                    .par_chunks(chunk_size)
                    .enumerate()
@@ -537,7 +537,7 @@ fn rayon_benches(c: &mut Criterion) {
                use rayon::prelude::*;
                let query_norm = vector_search::compute_norm(&query);
                let num_cores = rayon::current_num_threads().max(1);
-                let chunk_size = (n + num_cores - 1) / num_cores;
+                let chunk_size = n.div_ceil(num_cores);
                let mut results: Vec<(usize, f32)> = vectors
                    .par_chunks(chunk_size)
                    .enumerate()
@@ -766,12 +766,22 @@ fn adaptive_benches(c: &mut Criterion) {
        .map(|v| vector_search::compute_norm(v))
        .collect();
    let tombstones = vec![0u8; n];
+    let flat: Vec<f32> = vectors.iter().flatten().copied().collect();

    c.bench_function("adaptive_search_10k", |b| {
        let hw = HardwareCapabilities::detect();
        let strat = strategy::auto_select_strategy(n, &hw);
        b.iter(|| {
-            strategy::search_with_metrics(&query, &vectors, &norms, &tombstones, 10, strat, None)
+            strategy::search_with_metrics(
+                &query,
+                &vectors,
+                &flat,
+                &norms,
+                &tombstones,
+                10,
+                strat,
+                None,
+            )
        });
    });

@@ -781,6 +791,7 @@ fn adaptive_benches(c: &mut Criterion) {
            strategy::search_with_metrics(
                &query,
                &vectors,
+                &flat,
                &norms,
                &tombstones,
                10,
@@ -795,6 +806,7 @@ fn adaptive_benches(c: &mut Criterion) {
            strategy::search_with_metrics(
                &query,
                &vectors,
+                &flat,
                &norms,
                &tombstones,
                10,
@@ -809,6 +821,7 @@ fn adaptive_benches(c: &mut Criterion) {
            strategy::search_with_metrics(
                &query,
                &vectors,
+                &flat,
                &norms,
                &tombstones,
                10,
@@ -1,6 +1,7 @@
 use clawhdf5_agent::bm25::BM25Index;
 use clawhdf5_agent::consolidation::{
    ConsolidationConfig, ConsolidationEngine, ImportanceScorer, ImportanceWeights, MemorySource,
+    UntrustedSource,
 };
 use clawhdf5_agent::hybrid::{hybrid_search, rrf_hybrid_search};
 use clawhdf5_agent::knowledge::KnowledgeCache;
@@ -285,7 +286,12 @@ fn consolidation_benches(c: &mut Criterion) {
                        for i in 0..n {
                            let embedding = make_vec(&mut rng, DIM);
                            let chunk = format!("memory record {i} with some content");
-                            engine.add_memory(chunk, embedding, MemorySource::User, now + i as f64);
+                            engine.add_memory(
+                                chunk,
+                                embedding,
+                                UntrustedSource::User,
+                                now + i as f64,
+                            );
                        }
                        engine
                    },
@@ -307,9 +313,10 @@ fn consolidation_benches(c: &mut Criterion) {
        for i in 0..50usize {
            let embedding = make_vec(&mut rng, DIM);
            let chunk = format!("existing record {i}");
-            engine.add_memory(chunk, embedding, MemorySource::User, now + i as f64);
+            engine.add_memory(chunk, embedding, UntrustedSource::User, now + i as f64);
        }
        let records = engine.records().to_vec();
+        let record_refs: Vec<&_> = records.iter().collect();
        let weights = ImportanceWeights::default();
        let query_embedding = make_vec(&mut rng, DIM);
        let sample_text =
@@ -317,7 +324,7 @@ fn consolidation_benches(c: &mut Criterion) {

        group.bench_function("bench_importance_scoring", |b| {
            b.iter(|| {
-                let surprise = ImportanceScorer::score_surprise(&query_embedding, &records);
+                let surprise = ImportanceScorer::score_surprise(&query_embedding, &record_refs);
                let correction = ImportanceScorer::score_correction(&MemorySource::Correction);
                let length = ImportanceScorer::score_length(sample_text);
                ImportanceScorer::score_combined(surprise, correction, length, &weights)
@@ -354,7 +361,7 @@ fn temporal_benches(c: &mut Criterion) {
    // Insert benchmark: measure time to insert 10k timestamps one by one
    group.bench_function("bench_temporal_insert_10k", |b| {
        b.iter_batched(
-            || TemporalIndex::new(),
+            TemporalIndex::new,
            |mut idx| {
                for i in 0..N {
                    // Shuffle insertion order slightly using a simple offset pattern
@@ -442,7 +449,8 @@ fn large_consolidation_benches(c: &mut Criterion) {
    let mut group = c.benchmark_group("consolidation_large");
    group.sample_size(10);

-    for (label, n) in [("10k", 10_000usize)] {
+    {
+        let (label, n) = ("10k", 10_000usize);
        group.bench_with_input(
            BenchmarkId::new("bench_consolidation_cycle", label),
            &n,
@@ -459,7 +467,12 @@ fn large_consolidation_benches(c: &mut Criterion) {
                        for i in 0..n {
                            let embedding = make_vec(&mut rng, DIM);
                            let chunk = format!("memory record {i} with content");
-                            engine.add_memory(chunk, embedding, MemorySource::User, now + i as f64);
+                            engine.add_memory(
+                                chunk,
+                                embedding,
+                                UntrustedSource::User,
+                                now + i as f64,
+                            );
                        }
                        engine
                    },
@@ -0,0 +1,3 @@
+target/
+artifacts/
+coverage/
@@ -0,0 +1,23 @@
+[package]
+name = "clawhdf5-agent-fuzz"
+version = "0.0.0"
+publish = false
+edition = "2024"
+
+[package.metadata]
+cargo-fuzz = true
+
+[dependencies]
+libfuzzer-sys = "0.4"
+tempfile = "3"
+
+[dependencies.clawhdf5-agent]
+path = ".."
+
+[workspace]
+members = ["."]
+
+[[bin]]
+name = "fuzz_wal_replay"
+path = "fuzz_targets/fuzz_wal_replay.rs"
+doc = false
@@ -0,0 +1,36 @@
+#![no_main]
+//! Arbitrary bytes as a WAL file. Reading, and opening for append (which scans
+//! the chain and truncates an unverifiable tail), must never panic, hang, or
+//! allocate without bound — and after `open` repairs the file, everything
+//! `read_entries` returned before must still be returned.
+//!
+//! The deterministic counterpart that runs in ordinary CI is
+//! `tests/wal_properties.rs`; this target explores inputs it cannot reach.
+
+use std::io::Write as _;
+
+use clawhdf5_agent::wal::WalFile;
+use libfuzzer_sys::fuzz_target;
+
+fuzz_target!(|data: &[u8]| {
+    let Ok(mut tmp) = tempfile::NamedTempFile::new() else {
+        return;
+    };
+    if tmp.write_all(data).and_then(|()| tmp.flush()).is_err() {
+        return;
+    }
+    let before = WalFile::read_entries(tmp.path()).map(|e| e.len());
+    // Only the chained formats (header versions 3 and 4) are repaired in
+    // place. `open` deliberately recreates a legacy-format file from scratch:
+    // `HDF5Memory::open` has already replayed its entries by then.
+    let chained = matches!(data.get(4), Some(3 | 4));
+    let opened = WalFile::open(tmp.path());
+    if !chained {
+        return;
+    }
+    if let (Ok(before), Ok(wal)) = (before, opened) {
+        drop(wal);
+        let after = WalFile::read_entries(tmp.path()).map(|e| e.len());
+        assert_eq!(after.ok(), Some(before), "open() changed what is replayable");
+    }
+});
@@ -118,6 +118,7 @@ mod tests {
            created_at: "2025-01-01T00:00:00Z".to_string(),
            wal_enabled: false,
            wal_max_entries: 500,
+            quantized_index: false,
        }
    }

@@ -82,6 +82,68 @@ impl Default for AnomalyConfig {
    }
 }

+// ---------------------------------------------------------------------------
+// Pattern-match normalization
+// ---------------------------------------------------------------------------
+
+/// `true` for characters used to invisibly break up text without being
+/// rendered (zero-width joiners/spacers, bidi control marks, the BOM/ZWNBSP,
+/// soft hyphen, and the invisible math operators) — a common trick for
+/// splitting a flagged word so a literal-substring check misses it while the
+/// text still displays normally.
+fn is_invisible_format_char(ch: char) -> bool {
+    matches!(
+        ch,
+        '\u{00AD}' // soft hyphen
+            | '\u{200B}' // zero width space
+            | '\u{200C}' // zero width non-joiner
+            | '\u{200D}' // zero width joiner
+            | '\u{200E}' // left-to-right mark
+            | '\u{200F}' // right-to-left mark
+            | '\u{2060}' // word joiner
+            | '\u{2061}'..='\u{2064}' // invisible times/plus/separator/function application
+            | '\u{202A}'..='\u{202E}' // bidi embedding/override controls
+            | '\u{FEFF}' // BOM / zero width no-break space
+    )
+}
+
+/// Normalize text before suspicious-pattern matching so the cheapest evasion
+/// tricks — extra whitespace, zero-width characters, or punctuation spliced
+/// between letters (e.g. `"s.y.s.t.e.m"`) — don't defeat a literal-substring
+/// check. Lowercases, drops invisible-format and control characters, drops
+/// punctuation entirely (not just collapses it, so split words rejoin), and
+/// collapses whitespace runs to a single space.
+///
+/// Does not perform Unicode NFKC normalization or confusable/homoglyph
+/// folding (see [`WriteAnomalyDetector::check_pattern_anomaly`]).
+fn normalize_for_pattern_match(text: &str) -> String {
+    let mut out = String::with_capacity(text.len());
+    let mut last_was_space = true; // trims leading whitespace for free
+    for ch in text.chars() {
+        if ch.is_control() || is_invisible_format_char(ch) {
+            continue;
+        }
+        if ch.is_whitespace() {
+            if !last_was_space {
+                out.push(' ');
+                last_was_space = true;
+            }
+            continue;
+        }
+        if ch.is_ascii_punctuation() {
+            continue;
+        }
+        for lower in ch.to_lowercase() {
+            out.push(lower);
+        }
+        last_was_space = false;
+    }
+    while out.ends_with(' ') {
+        out.pop();
+    }
+    out
+}
+
 // ---------------------------------------------------------------------------
 // WriteEvent
 // ---------------------------------------------------------------------------
@@ -99,6 +161,9 @@ pub struct WriteEvent {
 // WriteAnomalyDetector
 // ---------------------------------------------------------------------------

+/// Upper bound on distinct session ids the detector tracks at once.
+const MAX_TRACKED_SESSIONS: usize = 4096;
+
 /// Tracks write events and raises alerts for suspicious behaviour.
 #[derive(Debug)]
 pub struct WriteAnomalyDetector {
@@ -127,6 +192,23 @@ impl WriteAnomalyDetector {
        if event.timestamp > self.last_timestamp {
            self.last_timestamp = event.timestamp;
        }
+        // Bound the per-session map: a long-lived process sees an unbounded
+        // number of distinct session ids. When it overflows, forget the
+        // sessions with the fewest writes (they are furthest from the limit
+        // this map exists to enforce); the current one is re-added below.
+        if self.session_counts.len() >= MAX_TRACKED_SESSIONS
+            && !self.session_counts.contains_key(&event.session_id)
+        {
+            let mut counts: Vec<u32> = self.session_counts.values().copied().collect();
+            let keep_from = counts.len() / 2;
+            counts.select_nth_unstable(keep_from);
+            let threshold = counts[keep_from];
+            self.session_counts.retain(|_, c| *c >= threshold);
+            if self.session_counts.len() >= MAX_TRACKED_SESSIONS {
+                // Every session had the same count: drop them all.
+                self.session_counts.clear();
+            }
+        }
        *self
            .session_counts
            .entry(event.session_id.clone())
@@ -146,6 +228,13 @@ impl WriteAnomalyDetector {
    /// Returns an alert if the number of writes in the last 60 seconds exceeds
    /// `config.max_writes_per_minute`, or if any session has exceeded
    /// `config.max_writes_per_session`.
+    ///
+    /// The 60-second window is a single shared window across all
+    /// sessions/sources, so when it trips the alert additionally names the
+    /// top-contributing session and source within that window — a session
+    /// can never account for more of the window than the aggregate count, so
+    /// this attributes the same trip to its actual offender rather than
+    /// reporting only the anonymous aggregate total.
    pub fn check_rate_anomaly(&self) -> Option<AnomalyAlert> {
        let recent = self.window.len() as u32;
        if recent > self.config.max_writes_per_minute {
@@ -156,11 +245,31 @@ impl WriteAnomalyDetector {
            } else {
                Severity::Medium
            };
+
+            let mut per_session: std::collections::HashMap<&str, u32> =
+                std::collections::HashMap::new();
+            // MemorySource isn't Eq/Hash, so key by its Display string instead.
+            let mut per_source: std::collections::HashMap<String, u32> =
+                std::collections::HashMap::new();
+            for e in &self.window {
+                *per_session.entry(e.session_id.as_str()).or_insert(0) += 1;
+                *per_source.entry(e.source.to_string()).or_insert(0) += 1;
+            }
+            let top_session = per_session.iter().max_by_key(|&(_, &c)| c);
+            let top_source = per_source.iter().max_by_key(|&(_, &c)| c);
+
+            let attribution = match (top_session, top_source) {
+                (Some((session, s_count)), Some((source, r_count))) => format!(
+                    "; top contributor: session '{session}' with {s_count} writes, \
+                     source {source} with {r_count} writes"
+                ),
+                _ => String::new(),
+            };
            return Some(AnomalyAlert {
                severity,
                message: format!(
-                    "Rate limit exceeded: {} writes in last 60s (max {})",
-                    recent, self.config.max_writes_per_minute
+                    "Rate limit exceeded: {} writes in last 60s (max {}){}",
+                    recent, self.config.max_writes_per_minute, attribution
                ),
                timestamp: self.last_timestamp,
            });
@@ -188,11 +297,24 @@ impl WriteAnomalyDetector {
    // -----------------------------------------------------------------------

    /// Returns an alert if `chunk` contains any of the configured suspicious
-    /// patterns (case-insensitive).
+    /// patterns, after normalizing both sides to defeat the cheapest evasion
+    /// tricks (case, extra whitespace, punctuation between letters,
+    /// zero-width/invisible-formatting characters).
+    ///
+    /// This does not perform Unicode NFKC normalization or confusable/
+    /// homoglyph folding (e.g. Cyrillic 'а' standing in for Latin 'a') —
+    /// that needs a per-codepoint confusable table (Unicode's
+    /// `confusables.txt`) beyond what's practical to hand-roll correctly,
+    /// and no such crate is a dependency of this crate today. A determined
+    /// attacker using homoglyphs can still evade these patterns.
    pub fn check_pattern_anomaly(&self, chunk: &str) -> Option<AnomalyAlert> {
-        let lower = chunk.to_lowercase();
+        let normalized = normalize_for_pattern_match(chunk);
        for pattern in &self.config.suspicious_patterns {
-            if lower.contains(pattern.as_str()) {
+            let normalized_pattern = normalize_for_pattern_match(pattern);
+            if normalized_pattern.is_empty() {
+                continue;
+            }
+            if normalized.contains(&normalized_pattern) {
                let severity = if pattern.contains("ignore") || pattern.contains("override") {
                    Severity::Critical
                } else if pattern.contains("system") || pattern.contains("jailbreak") {
@@ -327,6 +449,57 @@ mod tests {
        assert!(alert.unwrap().severity >= Severity::Medium);
    }

+    /// A single session dominating the shared 60s window must be named in
+    /// the alert, not just the anonymous aggregate count — this is the case
+    /// the separate cumulative max_writes_per_session check doesn't cover
+    /// (the window can trip before the session's lifetime total does).
+    #[test]
+    fn rate_anomaly_names_offending_session() {
+        let mut det = WriteAnomalyDetector::new(cfg());
+        for i in 0..11 {
+            det.record_write(event(
+                1.0 + i as f64 * 0.1,
+                "flood-session",
+                MemorySource::User,
+            ));
+        }
+        let alert = det.check_rate_anomaly().unwrap();
+        assert!(
+            alert.message.contains("flood-session"),
+            "expected the offending session to be named, got: {}",
+            alert.message
+        );
+    }
+
+    /// When many distinct sessions jointly trip the shared window, the top
+    /// contributor named must actually be the one with the most writes.
+    #[test]
+    fn rate_anomaly_attributes_top_contributor_among_many_sessions() {
+        let mut det = WriteAnomalyDetector::new(cfg());
+        // 5 sessions with 1 write each (below any per-session limit)...
+        for i in 0..5 {
+            det.record_write(event(
+                1.0 + i as f64 * 0.1,
+                "minor-session",
+                MemorySource::User,
+            ));
+        }
+        // ...plus one session responsible for the majority of the flood.
+        for i in 0..8 {
+            det.record_write(event(
+                2.0 + i as f64 * 0.1,
+                "major-session",
+                MemorySource::User,
+            ));
+        }
+        let alert = det.check_rate_anomaly().unwrap();
+        assert!(
+            alert.message.contains("major-session"),
+            "expected the top contributor to be named, got: {}",
+            alert.message
+        );
+    }
+
    #[test]
    fn rate_anomaly_critical_3x() {
        let mut det = WriteAnomalyDetector::new(cfg());
@@ -395,6 +568,71 @@ mod tests {
        assert!(alert.is_some());
    }

+    // --- Pattern-match evasion hardening ---
+
+    #[test]
+    fn pattern_defeats_extra_whitespace() {
+        let det = WriteAnomalyDetector::new(cfg());
+        let alert = det.check_pattern_anomaly("please   ignore    previous   instructions");
+        assert!(alert.is_some(), "extra whitespace must not defeat matching");
+    }
+
+    #[test]
+    fn pattern_defeats_punctuation_splicing() {
+        let det = WriteAnomalyDetector::new(cfg());
+        let alert = det.check_pattern_anomaly("i.g.n.o.r.e p-r-e-v-i-o-u-s instructions");
+        assert!(
+            alert.is_some(),
+            "punctuation spliced between letters must not defeat matching"
+        );
+    }
+
+    #[test]
+    fn pattern_defeats_zero_width_space() {
+        let det = WriteAnomalyDetector::new(cfg());
+        // Zero-width space (U+200B) inserted mid-word.
+        let chunk = "ign\u{200B}ore previ\u{200B}ous instructions";
+        let alert = det.check_pattern_anomaly(chunk);
+        assert!(
+            alert.is_some(),
+            "zero-width space injection must not defeat matching"
+        );
+    }
+
+    #[test]
+    fn pattern_defeats_zero_width_joiner_and_bom() {
+        let det = WriteAnomalyDetector::new(cfg());
+        let chunk = "jail\u{200D}break\u{FEFF} attempt";
+        let alert = det.check_pattern_anomaly(chunk);
+        assert!(
+            alert.is_some(),
+            "ZWJ/BOM injection must not defeat matching"
+        );
+    }
+
+    #[test]
+    fn pattern_still_clean_after_normalization() {
+        let det = WriteAnomalyDetector::new(cfg());
+        // Normalization must not introduce false positives on ordinary text
+        // that merely contains punctuation and extra whitespace.
+        let alert =
+            det.check_pattern_anomaly("Well,  I think... the weather is nice today, right?");
+        assert!(alert.is_none());
+    }
+
+    #[test]
+    fn normalize_for_pattern_match_examples() {
+        assert_eq!(
+            normalize_for_pattern_match("i.g.n.o.r.e p-r-e-v-i-o-u-s"),
+            "ignore previous"
+        );
+        assert_eq!(
+            normalize_for_pattern_match("ign\u{200B}ore   previous"),
+            "ignore previous"
+        );
+        assert_eq!(normalize_for_pattern_match("SYSTEM:"), "system");
+    }
+
    #[test]
    fn pattern_jailbreak() {
        let det = WriteAnomalyDetector::new(cfg());
@@ -37,7 +37,7 @@
 //! let mem = AsyncHDF5Memory::open_with(path, config).await?;
 //! mem.save(entry).await?;           // buffered → background writer
 //! mem.save_batch(entries).await?;    // also buffered
-//! let results = mem.hybrid_search(emb, "query".into(), 0.7, 0.3, 5).await;
+//! let results = mem.hybrid_search(emb, "query".into(), 0.4, 0.6, 5).await;
 //! mem.shutdown().await?;             // final flush + stop
 //! ```

@@ -408,6 +408,10 @@ impl AsyncHDF5Memory {
        let (tx, rx) = oneshot::channel();
        let _ = self.write_tx.send(WriteCmd::Shutdown(tx)).await;
        let _ = rx.await;
+        // The writer task has stopped, so nothing can write through this
+        // handle any more: release the single-writer lock now rather than at
+        // drop, so the store can be reopened while `self` is still in scope.
+        self.inner.lock().await.release_store_lock();
        Ok(())
    }
 }
@@ -3,12 +3,38 @@
 //! Provides a standard BM25 (Okapi BM25) implementation with an in-memory
 //! inverted index. Tombstoned documents are excluded from indexing and search.
 //!
-//! Optimizations:
-//! - Cached IDF scores (don't recompute per query)
-//! - Sorted posting lists by doc_id for cache-friendly access
-//! - Block-Max WAND early termination
+//! The index is **incremental**: [`BM25Index::add_document`] and
+//! [`BM25Index::remove_document`] keep it exactly equivalent to one built from
+//! scratch over the same live documents, so a store can maintain one index for
+//! its lifetime instead of re-tokenising the whole corpus per query. To make
+//! that possible IDF is computed at query time (it depends on the live
+//! document count) rather than cached at build time.
+//!
+//! - Posting lists sorted by doc id
+//! - Bounded-heap top-k; results ordered by score, then doc id (deterministic)

-use std::collections::HashMap;
+use std::cmp::Reverse;
+use std::collections::{BinaryHeap, HashMap};
+
+/// `f32` wrapper providing a total order (via `total_cmp`) so BM25 scores can
+/// be kept in a `BinaryHeap`. Scores are always finite in practice (no NaN
+/// inputs reach this path), so `total_cmp`'s NaN ordering is never exercised.
+#[derive(Debug, Clone, Copy, PartialEq)]
+struct HeapScore(f32);
+
+impl Eq for HeapScore {}
+
+impl PartialOrd for HeapScore {
+    fn partial_cmp(&self, other: &Self) -> Option<std::cmp::Ordering> {
+        Some(self.cmp(other))
+    }
+}
+
+impl Ord for HeapScore {
+    fn cmp(&self, other: &Self) -> std::cmp::Ordering {
+        self.0.total_cmp(&other.0)
+    }
+}

 /// Default BM25 term-frequency saturation parameter.
 const DEFAULT_K1: f32 = 1.2;
@@ -20,10 +46,11 @@ const DEFAULT_B: f32 = 0.75;
 pub struct BM25Index {
    /// Inverted index: token -> sorted list of (doc_id, term_frequency).
    inverted: HashMap<String, Vec<(usize, u32)>>,
-    /// Cached IDF scores per token.
-    idf_cache: HashMap<String, f32>,
    /// Number of tokens in each document (0 for tombstoned docs).
    doc_lengths: Vec<u32>,
+    /// Sum of `doc_lengths` over live documents (keeps `avg_dl` exact under
+    /// incremental updates).
+    total_length: u64,
    /// Average document length across non-tombstoned docs.
    avg_dl: f32,
    /// Number of non-tombstoned documents.
@@ -32,19 +59,27 @@ pub struct BM25Index {
    k1: f32,
    /// BM25 b parameter.
    b: f32,
+    /// Applied to every document and query token, so the two always agree.
+    filter: TokenFilter,
 }

 impl BM25Index {
    /// Build a BM25 index from a set of documents, excluding tombstoned entries.
    pub fn build(documents: &[String], tombstones: &[u8]) -> Self {
+        Self::build_with(documents, tombstones, TokenFilter::default())
+    }
+
+    /// [`BM25Index::build`] with the token filter chosen explicitly.
+    pub fn build_with(documents: &[String], tombstones: &[u8], filter: TokenFilter) -> Self {
        let mut index = Self {
            inverted: HashMap::new(),
-            idf_cache: HashMap::new(),
            doc_lengths: vec![0; documents.len()],
+            total_length: 0,
            avg_dl: 0.0,
            num_docs: 0,
            k1: DEFAULT_K1,
            b: DEFAULT_B,
+            filter,
        };
        index.index_documents(documents, tombstones);
        index
@@ -56,112 +91,165 @@ impl BM25Index {
    /// Uses Block-Max WAND for early termination when remaining documents
    /// cannot beat the current top-k threshold.
    pub fn search(&self, query: &str, k: usize) -> Vec<(usize, f32)> {
-        if self.num_docs == 0 || k == 0 {
+        if k == 0 {
            return Vec::new();
        }
-
-        let tokens = tokenize(query);
-        if tokens.is_empty() {
-            return Vec::new();
-        }
-
-        // Collect posting lists and cached IDF scores for query tokens
-        type QueryTerm<'a> = (&'a str, f32, &'a [(usize, u32)]);
-        let mut query_terms: Vec<QueryTerm<'_>> = Vec::new();
-        for token in &tokens {
-            if let (Some(postings), Some(&idf)) = (
-                self.inverted.get(token.as_str()),
-                self.idf_cache.get(token.as_str()),
-            ) {
-                query_terms.push((token, idf, postings));
+        // Top-k with a bounded min-heap: O(matches * log k) instead of sorting
+        // every match. Ties break towards the lower doc id so results are
+        // deterministic.
+        let mut heap: BinaryHeap<Reverse<(HeapScore, Reverse<usize>)>> =
+            BinaryHeap::with_capacity(k.min(1024) + 1);
+        for (doc_id, score) in self.scores(query) {
+            heap.push(Reverse((HeapScore(score), Reverse(doc_id))));
+            if heap.len() > k {
+                heap.pop();
            }
        }
+        let mut results: Vec<(usize, f32)> = heap
+            .into_iter()
+            .map(|Reverse((HeapScore(score), Reverse(doc_id)))| (doc_id, score))
+            .collect();
+        results.sort_by(|a, b| b.1.total_cmp(&a.1).then(a.0.cmp(&b.0)));
+        results
+    }

-        if query_terms.is_empty() {
+    /// The BM25 score of **every** matching document, in doc-id order, unsorted
+    /// by score. Score fusion normalises over the whole matching set, so it
+    /// needs all of these but not their ranking; producing a ranked list of
+    /// every match (`search(query, corpus_len)`) spent most of its time sorting.
+    pub fn scores(&self, query: &str) -> Vec<(usize, f32)> {
+        if self.num_docs == 0 {
            return Vec::new();
        }
-
-        // Accumulate BM25 scores per document using WAND-style scoring
-        let mut scores: HashMap<usize, f32> = HashMap::new();
-
-        // Compute maximum possible contribution per term for WAND
-        let max_tf_score: Vec<f32> = query_terms
-            .iter()
-            .map(|(_, idf, _)| {
-                // Upper bound: max TF contribution when tf is high and dl is short
-                let max_tf_num = 10.0 * (self.k1 + 1.0);
-                let max_tf_den = 10.0 + self.k1 * (1.0 - self.b);
-                idf * max_tf_num / max_tf_den
-            })
-            .collect();
-
-        let total_max_contribution: f32 = max_tf_score.iter().sum();
-
-        // Threshold for WAND early termination
-        let mut threshold = 0.0f32;
-        let mut top_k_scores: Vec<f32> = Vec::with_capacity(k);
-
-        for (term_idx, (_, idf, postings)) in query_terms.iter().enumerate() {
-            for &(doc_id, freq) in *postings {
+        // Term-at-a-time accumulation into a dense array: a common term has a
+        // posting per document, and hashing each one dominated query time.
+        // IDF is computed here rather than cached at build time: it depends on
+        // the live document count, which changes with every incremental
+        // add/remove, and costs one `ln` per query term.
+        let mut acc = vec![0.0f32; self.doc_lengths.len()];
+        let mut matched = false;
+        for token in tokenize_with(query, self.filter) {
+            let Some(postings) = self.inverted.get(token.as_str()) else {
+                continue;
+            };
+            matched = true;
+            let df = postings.len() as f32;
+            let idf = ((self.num_docs as f32 - df + 0.5) / (df + 0.5) + 1.0).ln();
+            for &(doc_id, freq) in postings {
                let dl = self.doc_lengths[doc_id] as f32;
                let freq_f = freq as f32;
                let tf = (freq_f * (self.k1 + 1.0))
                    / (freq_f + self.k1 * (1.0 - self.b + self.b * dl / self.avg_dl));
-                let contribution = idf * tf;
-
-                let entry = scores.entry(doc_id).or_insert(0.0);
-                *entry += contribution;
-
-                // WAND check: if this doc's current partial score + remaining
-                // max terms can't beat threshold, we can skip (but we still
-                // accumulate since we process term-at-a-time)
-                if term_idx == query_terms.len() - 1 {
-                    // Last term: check if this doc beats threshold
-                    let final_score = *entry;
-                    if final_score > threshold && top_k_scores.len() >= k {
-                        // Update threshold
-                        top_k_scores
-                            .sort_by(|a, b| b.partial_cmp(a).unwrap_or(std::cmp::Ordering::Equal));
-                        if final_score > top_k_scores[k - 1] {
-                            top_k_scores[k - 1] = final_score;
-                            top_k_scores.sort_by(|a, b| {
-                                b.partial_cmp(a).unwrap_or(std::cmp::Ordering::Equal)
-                            });
-                            threshold = top_k_scores[k - 1];
-                        }
-                    } else if top_k_scores.len() < k {
-                        top_k_scores.push(final_score);
-                        if top_k_scores.len() == k {
-                            top_k_scores.sort_by(|a, b| {
-                                b.partial_cmp(a).unwrap_or(std::cmp::Ordering::Equal)
-                            });
-                            threshold = top_k_scores[k - 1];
-                        }
-                    }
-                }
-            }
-            // After processing each term, check if remaining terms can
-            // possibly produce results above threshold
-            let remaining_max: f32 = max_tf_score[term_idx + 1..].iter().sum();
-            if remaining_max < threshold && total_max_contribution > 0.0 {
-                // Early termination: remaining terms can't produce new top-k
-                // entries on their own. But existing partial scores may still
-                // be updated, so we continue (WAND is approximate here).
-                let _ = remaining_max; // hint to compiler
+                acc[doc_id] += idf * tf;
            }
        }
+        if !matched {
+            return Vec::new();
+        }
+        // Every contribution is strictly positive (idf = ln(1 + x), x > 0), so
+        // a zero entry is a document no query term touched.
+        acc.into_iter()
+            .enumerate()
+            .filter(|&(_, score)| score > 0.0)
+            .collect()
+    }

-        let mut results: Vec<(usize, f32)> = scores.into_iter().collect();
-        results.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap_or(std::cmp::Ordering::Equal));
-        results.truncate(k);
-        results
+    /// The token filter this index was built with.
+    pub fn token_filter(&self) -> TokenFilter {
+        self.filter
+    }
+
+    /// Number of document slots (live or not) the index covers. Ids are
+    /// positions in the document list it mirrors.
+    pub fn len(&self) -> usize {
+        self.doc_lengths.len()
+    }
+
+    /// `true` when the index covers no document slots.
+    pub fn is_empty(&self) -> bool {
+        self.doc_lengths.is_empty()
+    }
+
+    /// Index `text` as document `doc_id`, which must be the next free id
+    /// (`self.len()`) or an existing slot that is currently empty (removed or
+    /// tombstoned). After any sequence of `add_document` / `remove_document`
+    /// calls the index scores exactly as one freshly built from the same live
+    /// documents.
+    pub fn add_document(&mut self, doc_id: usize, text: &str) {
+        if doc_id >= self.doc_lengths.len() {
+            self.doc_lengths.resize(doc_id + 1, 0);
+        }
+        debug_assert_eq!(self.doc_lengths[doc_id], 0, "slot {doc_id} is occupied");
+
+        let tokens = tokenize_with(text, self.filter);
+        let mut term_freqs: HashMap<&str, u32> = HashMap::new();
+        for token in &tokens {
+            *term_freqs.entry(token).or_insert(0) += 1;
+        }
+        for (token, freq) in term_freqs {
+            let postings = self.inverted.entry(token.to_string()).or_default();
+            // Posting lists stay sorted by doc id; appends are the common case.
+            match postings.last() {
+                Some(&(last, _)) if last >= doc_id => {
+                    let at = postings.partition_point(|&(id, _)| id < doc_id);
+                    postings.insert(at, (doc_id, freq));
+                }
+                _ => postings.push((doc_id, freq)),
+            }
+        }
+        self.doc_lengths[doc_id] = tokens.len() as u32;
+        self.total_length += tokens.len() as u64;
+        self.num_docs += 1;
+        self.refresh_avg_dl();
+    }
+
+    /// Extend the index to cover `len` document slots, leaving new ones empty.
+    /// Used for slots that hold no live document (tombstoned records).
+    pub fn pad_to(&mut self, len: usize) {
+        if len > self.doc_lengths.len() {
+            self.doc_lengths.resize(len, 0);
+        }
+    }
+
+    /// Remove document `doc_id`, whose indexed text was `text`. The text is
+    /// needed to find its postings; pass exactly what was added.
+    pub fn remove_document(&mut self, doc_id: usize, text: &str) {
+        let tokens = tokenize_with(text, self.filter);
+        let mut seen: std::collections::HashSet<&str> = std::collections::HashSet::new();
+        for token in &tokens {
+            if !seen.insert(token) {
+                continue;
+            }
+            if let Some(postings) = self.inverted.get_mut(token.as_str()) {
+                if let Ok(at) = postings.binary_search_by_key(&doc_id, |&(id, _)| id) {
+                    postings.remove(at);
+                }
+                if postings.is_empty() {
+                    self.inverted.remove(token.as_str());
+                }
+            }
+        }
+        if let Some(len) = self.doc_lengths.get_mut(doc_id) {
+            self.total_length = self.total_length.saturating_sub(u64::from(*len));
+            *len = 0;
+        }
+        self.num_docs = self.num_docs.saturating_sub(1);
+        self.refresh_avg_dl();
+    }
+
+    fn refresh_avg_dl(&mut self) {
+        self.avg_dl = if self.num_docs > 0 {
+            self.total_length as f32 / self.num_docs as f32
+        } else {
+            0.0
+        };
    }

    /// Rebuild the index from scratch (e.g., after compaction).
    pub fn rebuild(&mut self, documents: &[String], tombstones: &[u8]) {
        self.inverted.clear();
-        self.idf_cache.clear();
        self.doc_lengths = vec![0; documents.len()];
+        self.total_length = 0;
        self.avg_dl = 0.0;
        self.num_docs = 0;
        self.index_documents(documents, tombstones);
@@ -177,7 +265,7 @@ impl BM25Index {
                continue;
            }

-            let tokens = tokenize(doc);
+            let tokens = tokenize_with(doc, self.filter);
            let doc_len = tokens.len() as u32;
            self.doc_lengths[i] = doc_len;
            total_length += doc_len as u64;
@@ -198,33 +286,98 @@ impl BM25Index {
        }

        self.num_docs = count;
-        self.avg_dl = if count > 0 {
-            total_length as f32 / count as f32
-        } else {
-            0.0
-        };
+        self.total_length = total_length;
+        self.refresh_avg_dl();

        // Sort posting lists by doc_id for cache-friendly access
        for postings in self.inverted.values_mut() {
            postings.sort_by_key(|&(doc_id, _)| doc_id);
        }
-
-        // Pre-compute and cache IDF scores
-        for (token, postings) in &self.inverted {
-            let df = postings.len() as f32;
-            let idf = ((self.num_docs as f32 - df + 0.5) / (df + 0.5) + 1.0).ln();
-            self.idf_cache.insert(token.clone(), idf);
-        }
    }
 }

 /// Tokenize a string: lowercase, split on non-alphanumeric characters,
 /// filter empty tokens.
+/// What [`tokenize_with`] does to each token after splitting.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
+pub enum TokenFilter {
+    /// Lowercase and split only — the original behaviour.
+    #[default]
+    Plain,
+    /// Also strip common English inflections, so "running" and "runs" match
+    /// "run". Conservative on purpose: only plural and past/continuous verb
+    /// endings, and only on tokens long enough that stripping leaves a real
+    /// stem. A stemmer earns its keep by conflating *related* words; an
+    /// aggressive one also conflates unrelated ones ("universe"/"university"),
+    /// which costs precision.
+    Stemmed,
+}
+
+/// Strip common English inflections from an already-lowercased token.
+///
+/// Applied identically to documents and queries, so the pair only has to agree
+/// with itself — the stem need not be a real word.
+fn stem(token: &str) -> &str {
+    // Below this, stripping does more harm than good ("bed" -> "b").
+    const MIN_STEM: usize = 4;
+    let strip = |suffix: &str, min_len: usize| -> Option<&str> {
+        let stem = token.strip_suffix(suffix)?;
+        (stem.len() >= min_len).then_some(stem)
+    };
+
+    // Plurals first: "studies" -> "studi", "classes" -> "class", "cats" -> "cat".
+    // "ies" keeps its "i" so the result meets "-ied" ("studied" -> "studi").
+    if let Some(stem) = strip("ies", 2) {
+        return &token[..stem.len() + 1];
+    }
+    for suffix in ["sses", "shes", "ches", "xes", "zes"] {
+        if let Some(stem) = strip(suffix, MIN_STEM - 1) {
+            // Keep the sibilant: "classes" -> "class", not "clas".
+            return &token[..stem.len() + 2];
+        }
+    }
+    // Verb endings before the bare plural, so "raced" doesn't become "raced".
+    if let Some(stem) = strip("ing", MIN_STEM - 1).or_else(|| strip("ed", MIN_STEM - 1)) {
+        return undouble(stem);
+    }
+    if !token.ends_with("ss")
+        && !token.ends_with("us")
+        && !token.ends_with("is")
+        && let Some(stem) = strip("s", MIN_STEM - 1)
+    {
+        return stem;
+    }
+    token
+}
+
+/// "runn" -> "run": undo the consonant doubling that "-ing"/"-ed" introduce.
+fn undouble(stem: &str) -> &str {
+    let mut chars = stem.chars().rev();
+    let (Some(last), Some(prev)) = (chars.next(), chars.next()) else {
+        return stem;
+    };
+    let doubled = last == prev && !"aeiou".contains(last) && last.is_ascii_alphabetic();
+    if doubled && stem.len() > 3 {
+        &stem[..stem.len() - 1]
+    } else {
+        stem
+    }
+}
+
+#[cfg(test)]
 fn tokenize(text: &str) -> Vec<String> {
+    tokenize_with(text, TokenFilter::Plain)
+}
+
+/// Split `text` into scoring tokens under `filter`.
+pub fn tokenize_with(text: &str, filter: TokenFilter) -> Vec<String> {
    text.to_lowercase()
        .split(|c: char| !c.is_alphanumeric())
        .filter(|s| !s.is_empty())
-        .map(|s| s.to_string())
+        .map(|token| match filter {
+            TokenFilter::Plain => token.to_string(),
+            TokenFilter::Stemmed => stem(token).to_string(),
+        })
        .collect()
 }

@@ -370,24 +523,21 @@ mod tests {
    }

    #[test]
-    fn cached_idf_consistent_with_computed() {
+    fn score_matches_the_bm25_formula() {
        let docs = vec![
            "rust programming".to_string(),
            "rust systems".to_string(),
            "python scripting".to_string(),
        ];
-        let tombstones = vec![0, 0, 0];
-        let index = BM25Index::build(&docs, &tombstones);
+        let index = BM25Index::build(&docs, &[0, 0, 0]);

-        // IDF for "rust" (appears in 2 of 3 docs)
-        let idf_rust = index.idf_cache.get("rust").unwrap();
-        let expected_idf = ((3.0f32 - 2.0 + 0.5) / (2.0 + 0.5) + 1.0).ln();
-        assert!(
-            (idf_rust - expected_idf).abs() < 1e-6,
-            "cached IDF mismatch: {} vs {}",
-            idf_rust,
-            expected_idf
-        );
+        // "python": df = 1 of N = 3. Every doc has the average length (2) and
+        // tf = 1, so the tf factor is exactly 1 and the score is the IDF.
+        let results = index.search("python", 3);
+        let expected_idf = ((3.0f32 - 1.0 + 0.5) / (1.0 + 0.5) + 1.0).ln();
+        assert_eq!(results.len(), 1);
+        assert_eq!(results[0].0, 2);
+        assert!((results[0].1 - expected_idf).abs() < 1e-6, "{results:?}");
    }

    #[test]
@@ -451,4 +601,144 @@ mod tests {
            );
        }
    }
+    /// Documents drawn from a small vocabulary so terms collide heavily.
+    fn random_doc(state: &mut u64) -> String {
+        const VOCAB: &[&str] = &[
+            "alpha", "beta", "gamma", "delta", "eps", "zeta", "eta", "x1",
+        ];
+        let mut next = || {
+            *state = state
+                .wrapping_mul(6364136223846793005)
+                .wrapping_add(1442695040888963407);
+            (*state >> 33) as usize
+        };
+        let len = 1 + next() % 9;
+        (0..len)
+            .map(|_| VOCAB[next() % VOCAB.len()])
+            .collect::<Vec<_>>()
+            .join(" ")
+    }
+
+    #[test]
+    fn incremental_updates_match_a_fresh_build_exactly() {
+        for seed in 0..60u64 {
+            let mut state = seed.wrapping_mul(0x9E37_79B9_7F4A_7C15) | 1;
+            let mut docs: Vec<String> = Vec::new();
+            let mut tombstones: Vec<u8> = Vec::new();
+            let mut index = BM25Index::build(&docs, &tombstones);
+
+            for step in 0..80 {
+                state = state.wrapping_mul(6364136223846793005).wrapping_add(1);
+                let live: Vec<usize> = (0..docs.len()).filter(|&i| tombstones[i] == 0).collect();
+                match (state >> 40) % 4 {
+                    0 if !live.is_empty() => {
+                        // delete
+                        let id = live[(state >> 20) as usize % live.len()];
+                        index.remove_document(id, &docs[id]);
+                        tombstones[id] = 1;
+                    }
+                    1 if !live.is_empty() => {
+                        // update in place
+                        let id = live[(state >> 20) as usize % live.len()];
+                        let new_text = random_doc(&mut state);
+                        index.remove_document(id, &docs[id]);
+                        index.add_document(id, &new_text);
+                        docs[id] = new_text;
+                    }
+                    _ => {
+                        let text = random_doc(&mut state);
+                        index.add_document(docs.len(), &text);
+                        docs.push(text);
+                        tombstones.push(0);
+                    }
+                }
+
+                let fresh = BM25Index::build(&docs, &tombstones);
+                for query in ["alpha", "beta gamma", "x1 zeta alpha delta", "missing"] {
+                    let got = index.search(query, 5);
+                    let want = fresh.search(query, 5);
+                    assert_eq!(got.len(), want.len(), "seed {seed} step {step} {query:?}");
+                    for (g, w) in got.iter().zip(&want) {
+                        assert_eq!(
+                            g.0, w.0,
+                            "seed {seed} step {step} {query:?}: {got:?} vs {want:?}"
+                        );
+                        assert!(
+                            (g.1 - w.1).abs() < 1e-5,
+                            "seed {seed} step {step} {query:?}"
+                        );
+                    }
+                }
+            }
+        }
+    }
+
+    #[test]
+    fn scores_is_the_unranked_form_of_a_full_search() {
+        let mut state = 99u64;
+        let docs: Vec<String> = (0..200).map(|_| random_doc(&mut state)).collect();
+        let tombstones: Vec<u8> = (0..200).map(|i| u8::from(i % 7 == 0)).collect();
+        let index = BM25Index::build(&docs, &tombstones);
+        for query in ["alpha", "beta gamma x1", "missing", ""] {
+            let mut all = index.scores(query);
+            all.sort_by(|a, b| b.1.total_cmp(&a.1).then(a.0.cmp(&b.0)));
+            assert_eq!(all, index.search(query, docs.len()), "{query:?}");
+            assert!(all.iter().all(|(id, _)| tombstones[*id] == 0));
+        }
+    }
+
+    #[test]
+    fn stemming_conflates_inflections_of_the_same_word() {
+        let stem_of = |w: &str| tokenize_with(w, TokenFilter::Stemmed).pop().unwrap();
+        // Pairs that should meet.
+        for (a, b) in [
+            ("running", "runs"),
+            ("trained", "training"),
+            ("miles", "mile"),
+            ("studies", "studied"),
+            ("mentioned", "mentioning"),
+            ("classes", "class"),
+            ("planned", "planning"),
+        ] {
+            assert_eq!(stem_of(a), stem_of(b), "{a} / {b} should share a stem");
+        }
+        // Pairs that must stay apart. Note which pairs are deliberately absent:
+        // "bed"/"bedding" and "gas"/"gassed" both collapse to one stem, which
+        // is what Porter does too and is right — they are related words.
+        for (a, b) in [
+            ("universe", "university"),
+            ("business", "busy"),
+            ("this", "thing"),
+        ] {
+            assert_ne!(stem_of(a), stem_of(b), "{a} / {b} must not be conflated");
+        }
+        // Short words and non-inflections are left alone.
+        for word in ["run", "bus", "is", "his", "data", "gas"] {
+            assert_eq!(stem_of(word), word, "{word} should be untouched");
+        }
+    }
+
+    #[test]
+    fn stemming_is_off_by_default_and_applied_consistently() {
+        assert_eq!(tokenize("Running miles"), ["running", "miles"]);
+        assert_eq!(
+            tokenize_with("Running miles", TokenFilter::Stemmed),
+            ["run", "mile"]
+        );
+
+        // A query inflected differently from the document still matches.
+        let docs = vec!["I ran while training for the marathon".to_string()];
+        let plain = BM25Index::build_with(&docs, &[0], TokenFilter::Plain);
+        let stemmed = BM25Index::build_with(&docs, &[0], TokenFilter::Stemmed);
+        assert!(plain.search("trains", 1).is_empty());
+        assert_eq!(stemmed.search("trains", 1).len(), 1);
+    }
+
+    #[test]
+    fn ties_break_towards_the_lower_doc_id() {
+        let docs: Vec<String> = (0..6).map(|_| "same text".to_string()).collect();
+        let index = BM25Index::build(&docs, &[0; 6]);
+        let ids: Vec<usize> = index.search("same", 3).into_iter().map(|r| r.0).collect();
+        assert_eq!(ids, [0, 1, 2]);
+    }
 }
@@ -2,11 +2,143 @@

 use crate::vector_search;

+/// 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>>,
+    pub embeddings: Embeddings,
    pub source_channels: Vec<String>,
    pub timestamps: Vec<f64>,
    pub session_ids: Vec<String>,
@@ -23,7 +155,7 @@ impl MemoryCache {
    pub fn new(embedding_dim: usize) -> Self {
        Self {
            chunks: Vec::new(),
-            embeddings: Vec::new(),
+            embeddings: Embeddings::new(embedding_dim),
            source_channels: Vec::new(),
            timestamps: Vec::new(),
            session_ids: Vec::new(),
@@ -35,6 +167,16 @@ impl MemoryCache {
        }
    }

+    /// 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).
    pub fn len(&self) -> usize {
        self.chunks.len()
@@ -62,7 +204,7 @@ impl MemoryCache {
        let idx = self.chunks.len();
        let norm = vector_search::compute_norm(&embedding);
        self.chunks.push(chunk);
-        self.embeddings.push(embedding);
+        self.embeddings.push(&embedding);
        self.source_channels.push(source_channel);
        self.timestamps.push(timestamp);
        self.session_ids.push(session_id);
@@ -100,7 +242,7 @@ impl MemoryCache {
        if idx < self.chunks.len() {
            let norm = vector_search::compute_norm(&embedding);
            self.chunks[idx] = chunk;
-            self.embeddings[idx] = embedding;
+            self.embeddings.set(idx, &embedding);
            self.source_channels[idx] = source_channel;
            self.timestamps[idx] = timestamp;
            self.session_ids[idx] = session_id;
@@ -152,7 +294,7 @@ 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());
@@ -165,7 +307,8 @@ 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;
@@ -177,12 +320,123 @@ impl MemoryCache {
        (removed, index_map)
    }

-    /// Flatten all embeddings into a single Vec<f32> for HDF5 storage.
-    pub fn flat_embeddings(&self) -> Vec<f32> {
-        let mut flat = Vec::with_capacity(self.embeddings.len() * self.embedding_dim);
-        for emb in &self.embeddings {
-            flat.extend_from_slice(emb);
-        }
-        flat
+    /// 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()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// `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.as_flat(), expected);
+    }
+
+    #[test]
+    fn push_keeps_flat_buffer_in_sync() {
+        let mut cache = MemoryCache::new(3);
+        cache.push(
+            "a".into(),
+            vec![1.0, 2.0, 3.0],
+            "chan".into(),
+            0.0,
+            "s1".into(),
+            String::new(),
+        );
+        cache.push(
+            "b".into(),
+            vec![4.0, 5.0, 6.0],
+            "chan".into(),
+            1.0,
+            "s1".into(),
+            String::new(),
+        );
+        assert_flat_in_sync(&cache);
+        assert_eq!(
+            cache.embeddings.as_flat(),
+            vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0]
+        );
+    }
+
+    #[test]
+    fn update_keeps_flat_buffer_in_sync() {
+        let mut cache = MemoryCache::new(3);
+        cache.push(
+            "a".into(),
+            vec![1.0, 2.0, 3.0],
+            "chan".into(),
+            0.0,
+            "s1".into(),
+            String::new(),
+        );
+        cache.push(
+            "b".into(),
+            vec![4.0, 5.0, 6.0],
+            "chan".into(),
+            1.0,
+            "s1".into(),
+            String::new(),
+        );
+        cache.update(
+            0,
+            "a2".into(),
+            vec![7.0, 8.0, 9.0],
+            "chan".into(),
+            2.0,
+            "s1".into(),
+        );
+        assert_flat_in_sync(&cache);
+        assert_eq!(
+            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"
+        );
+    }
+
+    #[test]
+    fn compact_keeps_flat_buffer_in_sync() {
+        let mut cache = MemoryCache::new(2);
+        cache.push(
+            "a".into(),
+            vec![1.0, 1.0],
+            "chan".into(),
+            0.0,
+            "s1".into(),
+            String::new(),
+        );
+        cache.push(
+            "b".into(),
+            vec![2.0, 2.0],
+            "chan".into(),
+            1.0,
+            "s1".into(),
+            String::new(),
+        );
+        cache.push(
+            "c".into(),
+            vec![3.0, 3.0],
+            "chan".into(),
+            2.0,
+            "s1".into(),
+            String::new(),
+        );
+        cache.mark_deleted(1);
+        cache.compact();
+        assert_flat_in_sync(&cache);
+        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
+            .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]);
    }
 }
@@ -16,6 +16,55 @@ pub enum MemorySource {
    Correction,
 }

+/// Source classification for content whose true origin is *not*
+/// independently verified by the caller of [`ConsolidationEngine::add_memory`]
+/// — arbitrary text forwarded from a user, a tool's output, or a retrieval
+/// pipeline. This is the only source set `add_memory` accepts; it cannot
+/// claim the `System`/`Correction` importance boost (see [`TrustedSource`]
+/// and [`ConsolidationEngine::add_trusted_memory`]) — a caller passing
+/// through untrusted content has no way to self-report an elevated trust
+/// level through this entry point.
+#[derive(Clone, Debug, PartialEq)]
+pub enum UntrustedSource {
+    User,
+    Tool,
+    Retrieval,
+}
+
+impl From<UntrustedSource> for MemorySource {
+    fn from(s: UntrustedSource) -> Self {
+        match s {
+            UntrustedSource::User => MemorySource::User,
+            UntrustedSource::Tool => MemorySource::Tool,
+            UntrustedSource::Retrieval => MemorySource::Retrieval,
+        }
+    }
+}
+
+/// Source classification for content whose elevated trust level has been
+/// independently verified by the caller — e.g. the library's own
+/// system-generated text, or a caller that ran its own correction-cue
+/// detection (as `memory_strategy::SaveOnUserCorrection` does) rather than
+/// forwarding a caller-supplied label verbatim. `MemorySource::System`/
+/// `Correction` get elevated importance weighting in
+/// [`ImportanceScorer::score_correction`]; only reachable through
+/// [`ConsolidationEngine::add_trusted_memory`], a distinct entry point from
+/// the one untrusted content is passed through.
+#[derive(Clone, Debug, PartialEq)]
+pub enum TrustedSource {
+    System,
+    Correction,
+}
+
+impl From<TrustedSource> for MemorySource {
+    fn from(s: TrustedSource) -> Self {
+        match s {
+            TrustedSource::System => MemorySource::System,
+            TrustedSource::Correction => MemorySource::Correction,
+        }
+    }
+}
+
 #[derive(Clone, Debug, PartialEq)]
 pub enum MemoryTier {
    Working,
@@ -118,7 +167,7 @@ impl ImportanceScorer {

    /// Novelty score: 1.0 − max cosine similarity against all existing records.
    /// Returns 1.0 when there are no existing memories.
-    pub fn score_surprise(embedding: &[f32], existing_memories: &[MemoryRecord]) -> f32 {
+    pub fn score_surprise(embedding: &[f32], existing_memories: &[&MemoryRecord]) -> f32 {
        if existing_memories.is_empty() {
            return 1.0;
        }
@@ -199,21 +248,51 @@ impl ConsolidationEngine {
        }
    }

-    /// Add a new memory to the Working tier.
+    /// Add a new memory to the Working tier from an untrusted/ordinary origin
+    /// (User, Tool, or Retrieval). This is the entry point for arbitrary
+    /// caller-supplied content — it cannot claim the elevated System/
+    /// Correction importance boost. Use [`Self::add_trusted_memory`] for
+    /// content whose elevated trust level the caller has independently
+    /// verified.
    ///
    /// Importance is scored against existing Working-tier records only.
    pub fn add_memory(
+        &mut self,
+        chunk: String,
+        embedding: Vec<f32>,
+        source: UntrustedSource,
+        now: f64,
+    ) -> u64 {
+        self.add_memory_with_source(chunk, embedding, source.into(), now)
+    }
+
+    /// Add a new memory tagged System or Correction, which get elevated
+    /// importance weighting in [`ImportanceScorer::score_correction`]. Only
+    /// call this from code that has independently verified the origin (the
+    /// library's own system-generated text, or a caller that ran its own
+    /// correction-cue detection) — never from a path that forwards a
+    /// caller-supplied trust label verbatim.
+    pub fn add_trusted_memory(
+        &mut self,
+        chunk: String,
+        embedding: Vec<f32>,
+        source: TrustedSource,
+        now: f64,
+    ) -> u64 {
+        self.add_memory_with_source(chunk, embedding, source.into(), now)
+    }
+
+    fn add_memory_with_source(
        &mut self,
        chunk: String,
        embedding: Vec<f32>,
        source: MemorySource,
        now: f64,
    ) -> u64 {
-        let working: Vec<MemoryRecord> = self
+        let working: Vec<&MemoryRecord> = self
            .records
            .iter()
            .filter(|r| r.tier == MemoryTier::Working)
-            .cloned()
            .collect();

        let surprise = ImportanceScorer::score_surprise(&embedding, &working);
@@ -281,7 +360,7 @@ impl ConsolidationEngine {
        if working_count > capacity {
            let evict_n = working_count - capacity;
            // Collect the ids of the records to evict (lowest decay = first in sorted list).
-            let evict_ids: Vec<u64> = working_indices[..evict_n]
+            let evict_ids: std::collections::HashSet<u64> = working_indices[..evict_n]
                .iter()
                .map(|&i| self.records[i].id)
                .collect();
@@ -342,7 +421,7 @@ impl ConsolidationEngine {
            });

            let evict_n = episodic_count - episodic_capacity;
-            let evict_ids: Vec<u64> = episodic_indices[..evict_n]
+            let evict_ids: std::collections::HashSet<u64> = episodic_indices[..evict_n]
                .iter()
                .map(|&i| self.records[i].id)
                .collect();
@@ -419,13 +498,44 @@ mod tests {
    // ---------------------------------------------------------------------------
    // 2. Add memory — basic
    // ---------------------------------------------------------------------------
+    /// add_trusted_memory(TrustedSource::Correction) must actually produce a
+    /// MemorySource::Correction record — the only way to reach that elevated
+    /// classification, since add_memory's UntrustedSource has no such variant.
+    #[test]
+    fn test_add_trusted_memory_sets_correction_source() {
+        let mut engine = ConsolidationEngine::new(ConsolidationConfig::default());
+        let id = engine.add_trusted_memory(
+            "verified correction".to_string(),
+            unit_vec(4, 0),
+            TrustedSource::Correction,
+            0.0,
+        );
+        let rec = engine.get_by_id(id).unwrap();
+        assert_eq!(rec.source, MemorySource::Correction);
+    }
+
+    /// add_trusted_memory(TrustedSource::System) must produce a
+    /// MemorySource::System record.
+    #[test]
+    fn test_add_trusted_memory_sets_system_source() {
+        let mut engine = ConsolidationEngine::new(ConsolidationConfig::default());
+        let id = engine.add_trusted_memory(
+            "bootstrap text".to_string(),
+            unit_vec(4, 0),
+            TrustedSource::System,
+            0.0,
+        );
+        let rec = engine.get_by_id(id).unwrap();
+        assert_eq!(rec.source, MemorySource::System);
+    }
+
    #[test]
    fn test_add_memory_basic() {
        let mut engine = ConsolidationEngine::new(ConsolidationConfig::default());
        let id = engine.add_memory(
            "Hello world".to_string(),
            unit_vec(4, 0),
-            MemorySource::User,
+            UntrustedSource::User,
            1_000_000.0,
        );
        assert_eq!(id, 0);
@@ -453,7 +563,7 @@ mod tests {
    #[test]
    fn test_importance_scorer_surprise_identical() {
        let emb = unit_vec(4, 0);
-        let existing = vec![MemoryRecord {
+        let existing = [MemoryRecord {
            id: 0,
            chunk: "existing".to_string(),
            embedding: emb.clone(),
@@ -464,7 +574,8 @@ mod tests {
            created_at: 0.0,
            source: MemorySource::User,
        }];
-        let score = ImportanceScorer::score_surprise(&emb, &existing);
+        let existing_refs: Vec<&MemoryRecord> = existing.iter().collect();
+        let score = ImportanceScorer::score_surprise(&emb, &existing_refs);
        assert!(score < 0.01, "expected ~0.0, got {score}");
    }

@@ -492,23 +603,20 @@ mod tests {
    fn test_importance_scorer_length() {
        assert!((ImportanceScorer::score_length("")).abs() < f32::EPSILON);
        // 50 words → 0.5
-        let fifty_words = std::iter::repeat("word")
-            .take(50)
+        let fifty_words = std::iter::repeat_n("word", 50)
            .collect::<Vec<_>>()
            .join(" ");
        let s50 = ImportanceScorer::score_length(&fifty_words);
        assert!((s50 - 0.5).abs() < 1e-5, "expected 0.5, got {s50}");

        // 100 words → 1.0
-        let hundred_words = std::iter::repeat("word")
-            .take(100)
+        let hundred_words = std::iter::repeat_n("word", 100)
            .collect::<Vec<_>>()
            .join(" ");
        assert_eq!(ImportanceScorer::score_length(&hundred_words), 1.0);

        // 200 words → still 1.0 (clamped)
-        let two_hundred = std::iter::repeat("word")
-            .take(200)
+        let two_hundred = std::iter::repeat_n("word", 200)
            .collect::<Vec<_>>()
            .join(" ");
        assert_eq!(ImportanceScorer::score_length(&two_hundred), 1.0);
@@ -582,9 +690,11 @@ mod tests {
    // ---------------------------------------------------------------------------
    #[test]
    fn test_consolidate_eviction_working() {
-        let mut cfg = ConsolidationConfig::default();
-        cfg.working_capacity = 3;
-        cfg.working_to_episodic_threshold = 2.0; // never promote in this test
+        let cfg = ConsolidationConfig {
+            working_capacity: 3,
+            working_to_episodic_threshold: 2.0, // never promote in this test
+            ..Default::default()
+        };
        let mut engine = ConsolidationEngine::new(cfg);

        // Add 5 records; all have very low importance so none get promoted.
@@ -592,7 +702,7 @@ mod tests {
            let id = engine.add_memory(
                "x".to_string(),
                unit_vec(4, i as usize),
-                MemorySource::User,
+                UntrustedSource::User,
                i as f64,
            );
            // Force low importance so promotion threshold is not crossed.
@@ -625,10 +735,10 @@ mod tests {
        let cfg = ConsolidationConfig::default();
        let mut engine = ConsolidationEngine::new(cfg);

-        let id = engine.add_memory(
+        let id = engine.add_trusted_memory(
            "important memory".to_string(),
            unit_vec(4, 0),
-            MemorySource::Correction,
+            TrustedSource::Correction,
            0.0,
        );
        // Force importance above threshold.
@@ -661,7 +771,7 @@ mod tests {
        let id = engine.add_memory(
            "frequently accessed".to_string(),
            unit_vec(4, 0),
-            MemorySource::User,
+            UntrustedSource::User,
            0.0,
        );

@@ -689,7 +799,12 @@ mod tests {
    #[test]
    fn test_access_memory_reactivation() {
        let mut engine = ConsolidationEngine::new(ConsolidationConfig::default());
-        let id = engine.add_memory("chunk".to_string(), unit_vec(4, 0), MemorySource::User, 0.0);
+        let id = engine.add_memory(
+            "chunk".to_string(),
+            unit_vec(4, 0),
+            UntrustedSource::User,
+            0.0,
+        );

        engine.access_memory(id, 5000.0);
        let rec = engine.get_by_id(id).unwrap();
@@ -710,11 +825,11 @@ mod tests {
        let mut engine = ConsolidationEngine::new(ConsolidationConfig::default());

        // 2 Working
-        engine.add_memory("w1".to_string(), unit_vec(4, 0), MemorySource::User, 0.0);
-        engine.add_memory("w2".to_string(), unit_vec(4, 1), MemorySource::User, 0.0);
+        engine.add_memory("w1".to_string(), unit_vec(4, 0), UntrustedSource::User, 0.0);
+        engine.add_memory("w2".to_string(), unit_vec(4, 1), UntrustedSource::User, 0.0);

        // 1 Episodic (manually set)
-        let id_e = engine.add_memory("e1".to_string(), unit_vec(4, 2), MemorySource::User, 0.0);
+        let id_e = engine.add_memory("e1".to_string(), unit_vec(4, 2), UntrustedSource::User, 0.0);
        engine
            .records
            .iter_mut()
@@ -723,7 +838,7 @@ mod tests {
            .tier = MemoryTier::Episodic;

        // 1 Semantic (manually set)
-        let id_s = engine.add_memory("s1".to_string(), unit_vec(4, 3), MemorySource::User, 0.0);
+        let id_s = engine.add_memory("s1".to_string(), unit_vec(4, 3), UntrustedSource::User, 0.0);
        engine
            .records
            .iter_mut()
@@ -742,9 +857,11 @@ mod tests {
    // ---------------------------------------------------------------------------
    #[test]
    fn test_consolidate_episodic_eviction() {
-        let mut cfg = ConsolidationConfig::default();
-        cfg.episodic_capacity = 3;
-        cfg.working_to_episodic_threshold = 2.0; // never auto-promote from Working
+        let cfg = ConsolidationConfig {
+            episodic_capacity: 3,
+            working_to_episodic_threshold: 2.0, // never auto-promote from Working
+            ..Default::default()
+        };
        let mut engine = ConsolidationEngine::new(cfg);

        // Seed 5 records directly in Episodic.
@@ -752,7 +869,7 @@ mod tests {
            let id = engine.add_memory(
                "episodic chunk".to_string(),
                unit_vec(4, i as usize),
-                MemorySource::User,
+                UntrustedSource::User,
                i as f64,
            );
            let rec = engine.records.iter_mut().find(|r| r.id == id).unwrap();
@@ -777,8 +777,10 @@ mod tests {

    #[test]
    fn test_tech_disabled() {
-        let mut config = ExtractorConfig::default();
-        config.extract_technology = false;
+        let config = ExtractorConfig {
+            extract_technology: false,
+            ..Default::default()
+        };
        let e = EntityExtractor::new(config);
        let entities = e.extract("We use Rust and Docker.");
        assert!(
@@ -847,8 +849,10 @@ mod tests {

    #[test]
    fn test_date_disabled() {
-        let mut config = ExtractorConfig::default();
-        config.extract_dates = false;
+        let config = ExtractorConfig {
+            extract_dates: false,
+            ..Default::default()
+        };
        let e = EntityExtractor::new(config);
        let entities = e.extract("Released on 2024-03-19.");
        assert!(
@@ -981,8 +985,10 @@ mod tests {

    #[test]
    fn test_confidence_filter() {
-        let mut config = ExtractorConfig::default();
-        config.min_confidence = 0.95;
+        let config = ExtractorConfig {
+            min_confidence: 0.95,
+            ..Default::default()
+        };
        let e = EntityExtractor::new(config);
        // Only dates (0.95) and techs (0.9) should survive; 0.9 < 0.95 filters techs.
        let entities = e.extract("We use Rust since 2024-01-01.");
@@ -1002,7 +1008,7 @@ mod tests {
    fn test_batch_dedup() {
        let e = default_extractor();
        let texts = ["We use Rust.", "Rust is fast.", "Also Rust for safety."];
-        let entities = e.extract_batch(&texts.iter().map(|s| *s).collect::<Vec<_>>());
+        let entities = e.extract_batch(&texts);
        let rust_count = entities.iter().filter(|x| x.text == "Rust").count();
        assert_eq!(rust_count, 1, "Rust should appear exactly once after dedup");
    }
@@ -1011,7 +1017,7 @@ mod tests {
    fn test_batch_multiple_types() {
        let e = default_extractor();
        let texts = ["Deploy with Docker.", "We merged last week."];
-        let entities = e.extract_batch(&texts.iter().map(|s| *s).collect::<Vec<_>>());
+        let entities = e.extract_batch(&texts);
        assert!(
            entities
                .iter()
@@ -28,13 +28,40 @@ use crate::vector_search;
 pub fn hybrid_search(
    query_embedding: &[f32],
    query_text: &str,
-    vectors: &[Vec<f32>],
-    _chunks: &[String],
+    vectors: &(impl crate::vector_search::VectorSet + Sync + ?Sized),
+    chunks: &[String],
    tombstones: &[u8],
    bm25_index: &BM25Index,
    vector_weight: f32,
    keyword_weight: f32,
    k: usize,
+) -> Vec<(usize, f32)> {
+    hybrid_search_fused(
+        query_embedding,
+        query_text,
+        vectors,
+        chunks,
+        tombstones,
+        bm25_index,
+        Fusion::Weighted {
+            vector: vector_weight,
+            keyword: keyword_weight,
+        },
+        k,
+    )
+}
+
+/// [`hybrid_search`] with the fusion method chosen explicitly.
+#[allow(clippy::too_many_arguments)]
+pub fn hybrid_search_fused(
+    query_embedding: &[f32],
+    query_text: &str,
+    vectors: &(impl crate::vector_search::VectorSet + Sync + ?Sized),
+    _chunks: &[String],
+    tombstones: &[u8],
+    bm25_index: &BM25Index,
+    fusion: Fusion,
+    k: usize,
 ) -> Vec<(usize, f32)> {
    // Get raw scores from both systems. Request all results so normalization
    // covers the full distribution.
@@ -42,12 +69,12 @@ pub fn hybrid_search(
    let 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)
@@ -58,9 +85,9 @@ pub fn hybrid_search(
            vector_search::cosine_similarity_batch(query_embedding, vectors, tombstones)
        }
    };
-    let kw_scores = bm25_index.search(query_text, vectors.len());
+    let kw_scores = bm25_index.scores(query_text);

-    merge_vector_keyword(vec_scores, kw_scores, vector_weight, keyword_weight, k)
+    fuse(vec_scores, kw_scores, fusion, k)
 }

 /// Merge pre-computed vector-similarity and keyword scores into a single ranking.
@@ -76,29 +103,120 @@ pub fn merge_vector_keyword(
    keyword_weight: f32,
    k: usize,
 ) -> Vec<(usize, f32)> {
-    // Normalize each set to [0, 1].
-    let vec_normalized = normalize_scores(&vec_scores);
-    let kw_normalized = normalize_scores(&kw_scores);
+    fuse(
+        vec_scores,
+        kw_scores,
+        Fusion::Weighted {
+            vector: vector_weight,
+            keyword: keyword_weight,
+        },
+        k,
+    )
+}

-    // Merge scores with weights.
-    let mut merged: HashMap<usize, f32> = HashMap::new();
+/// How the vector and keyword stages are combined into one ranking.
+#[derive(Debug, Clone, Copy, PartialEq)]
+pub enum Fusion {
+    /// Min-max normalise each stage over its own candidates, then take a
+    /// weighted sum. Uses the *scores*, so a stage that separates its
+    /// candidates sharply keeps that separation — and a stage whose candidates
+    /// are all near-identical contributes little.
+    Weighted {
+        /// Weight on the vector stage.
+        vector: f32,
+        /// Weight on the keyword stage.
+        keyword: f32,
+    },
+    /// Reciprocal rank fusion: each stage contributes `1 / (k + rank)`,
+    /// ignoring score magnitudes entirely. Robust when the two stages'
+    /// scores aren't comparable, at the cost of discarding confidence.
+    Rrf {
+        /// The rank-damping constant; 60 is the value from the original paper.
+        k: f32,
+    },
+}

-    for (idx, score) in &vec_normalized {
-        *merged.entry(*idx).or_insert(0.0) += vector_weight * score;
+impl Default for Fusion {
+    fn default() -> Self {
+        DEFAULT_FUSION
    }
-    for (idx, score) in &kw_normalized {
-        *merged.entry(*idx).or_insert(0.0) += keyword_weight * score;
+}
+
+/// The fusion `hybrid_search` uses unless told otherwise.
+///
+/// The weights are not a guess: a sweep of every 0.1 step over the full
+/// LongMemEval haystack (500 questions, real MiniLM embeddings) found the
+/// long-standing 0.7/0.3 default *strictly dominated* — 0.4/0.6 is better at
+/// Hit@1, Hit@5, Hit@10 and MRR, at both turn and session granularity. See
+/// `BENCHMARKS.md`, "Weight sweep".
+pub const DEFAULT_FUSION: Fusion = Fusion::Weighted {
+    vector: 0.4,
+    keyword: 0.6,
+};
+
+/// Combine one ranked candidate list from each stage into a single top-`k`.
+///
+/// Neither list need be sorted; both are consumed.
+pub fn fuse(
+    vec_scores: Vec<(usize, f32)>,
+    kw_scores: Vec<(usize, f32)>,
+    fusion: Fusion,
+    k: usize,
+) -> Vec<(usize, f32)> {
+    let mut merged: HashMap<usize, f32> = HashMap::new();
+    match fusion {
+        Fusion::Weighted { vector, keyword } => {
+            // Normalize each set to [0, 1].
+            for (idx, score) in &normalize_scores(&vec_scores) {
+                *merged.entry(*idx).or_insert(0.0) += vector * score;
+            }
+            for (idx, score) in &normalize_scores(&kw_scores) {
+                *merged.entry(*idx).or_insert(0.0) += keyword * score;
+            }
+        }
+        Fusion::Rrf { k: damping } => {
+            for mut stage in [vec_scores, kw_scores] {
+                // Rank 1 is the best score. Ties break by index so a stage's
+                // contribution doesn't depend on the candidate order it
+                // happened to be produced in.
+                stage.sort_by(|a, b| {
+                    b.1.partial_cmp(&a.1)
+                        .unwrap_or(std::cmp::Ordering::Equal)
+                        .then(a.0.cmp(&b.0))
+                });
+                for (rank, (idx, _)) in stage.iter().enumerate() {
+                    *merged.entry(*idx).or_insert(0.0) += 1.0 / (damping + (rank + 1) as f32);
+                }
+            }
+        }
    }

    let mut results: Vec<(usize, f32)> = merged.into_iter().collect();
-    results.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap_or(std::cmp::Ordering::Equal));
-    results.truncate(k);
+    // Index tie-break: `merged` is a HashMap, so without it the ties that
+    // survive differ from run to run.
+    let by_score_then_id = |a: &(usize, f32), b: &(usize, f32)| {
+        b.1.partial_cmp(&a.1)
+            .unwrap_or(std::cmp::Ordering::Equal)
+            .then(a.0.cmp(&b.0))
+    };
+    // Only the top k are wanted: partition them out, then order just those,
+    // instead of sorting every candidate (the keyword side can be the corpus).
+    if k == 0 {
+        return Vec::new();
+    }
+    if results.len() > k {
+        results.select_nth_unstable_by(k - 1, by_score_then_id);
+        results.truncate(k);
+    }
+    results.sort_by(by_score_then_id);
    results
 }

 /// Normalize a set of scores to the [0, 1] range using min-max normalization.
 ///
-/// If all scores are identical, returns 0.0 for each entry.
+/// If all scores are identical there is no spread to normalise: each entry
+/// gets 1.0 when that score is positive (all equally the best match) and 0.0
+/// otherwise (nothing matched).
 fn normalize_scores(scores: &[(usize, f32)]) -> Vec<(usize, f32)> {
    if scores.is_empty() {
        return Vec::new();
@@ -112,7 +230,13 @@ fn normalize_scores(scores: &[(usize, f32)]) -> Vec<(usize, f32)> {

    let range = max - min;
    if range == 0.0 {
-        return scores.iter().map(|(idx, _)| (*idx, 0.0)).collect();
+        // All candidates scored the same (including the single-candidate
+        // case), so min-max has no spread to work with. They are all equally
+        // the best match if that score is positive, and all non-matches
+        // otherwise. This used to return 0.0 unconditionally, which erased a
+        // lone perfect match from the fused score.
+        let level = if max > 0.0 { 1.0 } else { 0.0 };
+        return scores.iter().map(|(idx, _)| (*idx, level)).collect();
    }

    scores
@@ -146,7 +270,7 @@ 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,
@@ -158,12 +282,12 @@ 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)
@@ -174,7 +298,7 @@ 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));
@@ -324,10 +448,80 @@ mod tests {

    #[test]
    fn normalize_scores_single() {
+        // A lone positive score is the best match there is, not a non-match.
        let result = normalize_scores(&[(0, 5.0)]);
        assert_eq!(result.len(), 1);
-        // Single score normalizes to 0.0 (range is 0)
-        assert_eq!(result[0].1, 0.0);
+        assert_eq!(result[0].1, 1.0);
+    }
+
+    #[test]
+    fn default_fusion_is_the_tuned_operating_point() {
+        // A sweep over the full LongMemEval haystack found 0.7/0.3 strictly
+        // dominated by 0.4/0.6 (BENCHMARKS.md). This guards the finding
+        // against being quietly undone.
+        assert_eq!(
+            DEFAULT_FUSION,
+            Fusion::Weighted {
+                vector: 0.4,
+                keyword: 0.6
+            }
+        );
+    }
+
+    #[test]
+    fn rrf_rewards_agreement_between_the_stages_and_ignores_magnitudes() {
+        // Doc 1 is second-best in both stages; doc 0 is best in one and absent
+        // from the other. RRF prefers the doc both stages liked.
+        let vec_scores = vec![(0, 100.0), (1, 0.9)];
+        let kw_scores = vec![(2, 5.0), (1, 4.9)];
+        let ranked = fuse(vec_scores, kw_scores, Fusion::Rrf { k: 60.0 }, 3);
+        assert_eq!(ranked[0].0, 1, "{ranked:?}");
+
+        // Scaling one stage's scores cannot change an RRF ranking, only the
+        // order within that stage can.
+        let a = fuse(
+            vec![(0, 1.0), (1, 0.5)],
+            vec![(1, 2.0), (0, 1.0)],
+            Fusion::Rrf { k: 60.0 },
+            2,
+        );
+        let b = fuse(
+            vec![(0, 1e6), (1, -3.0)],
+            vec![(1, 0.002), (0, 0.001)],
+            Fusion::Rrf { k: 60.0 },
+            2,
+        );
+        assert_eq!(
+            a.iter().map(|r| r.0).collect::<Vec<_>>(),
+            b.iter().map(|r| r.0).collect::<Vec<_>>()
+        );
+    }
+
+    #[test]
+    fn merge_top_k_matches_a_full_sort() {
+        // Many ties (scores repeat) so the index tie-break is exercised.
+        let vec_scores: Vec<(usize, f32)> = (0..300).map(|i| (i, ((i * 7) % 13) as f32)).collect();
+        let kw_scores: Vec<(usize, f32)> = (100..500).map(|i| (i, ((i * 5) % 11) as f32)).collect();
+        let everything =
+            merge_vector_keyword(vec_scores.clone(), kw_scores.clone(), 0.7, 0.3, 10_000);
+        assert_eq!(everything.len(), 500);
+        assert!(
+            everything
+                .windows(2)
+                .all(|w| { w[0].1 > w[1].1 || (w[0].1 == w[1].1 && w[0].0 < w[1].0) })
+        );
+        for k in [0, 1, 7, 50, 499, 500, 501] {
+            let top = merge_vector_keyword(vec_scores.clone(), kw_scores.clone(), 0.7, 0.3, k);
+            assert_eq!(top, everything[..k.min(500)], "k = {k}");
+        }
+    }
+
+    #[test]
+    fn normalize_scores_all_equal() {
+        let matched = normalize_scores(&[(0, 0.4), (1, 0.4)]);
+        assert!(matched.iter().all(|(_, s)| *s == 1.0));
+        let unmatched = normalize_scores(&[(0, 0.0), (1, 0.0)]);
+        assert!(unmatched.iter().all(|(_, s)| *s == 0.0));
    }

    #[test]
@@ -50,6 +50,9 @@ impl RelationType {
 pub struct Entity {
    pub id: u64,
    pub name: String,
+    /// Lowercased `name`, cached at construction time to avoid re-allocating
+    /// and re-lowercasing on every entity-resolution scan.
+    pub name_lower: String,
    pub entity_type: String,
    /// Index into the memory embeddings array, or -1 if none.
    pub embedding_idx: i64,
@@ -69,6 +72,7 @@ impl Default for Entity {
        Self {
            id: 0,
            name: String::new(),
+            name_lower: String::new(),
            entity_type: String::new(),
            embedding_idx: -1,
            properties: HashMap::new(),
@@ -151,6 +155,55 @@ fn levenshtein(a: &str, b: &str) -> usize {
    prev[nb]
 }

+// ---------------------------------------------------------------------------
+// AdjacencyIndex
+// ---------------------------------------------------------------------------
+
+/// Adjacency index over a snapshot of `entities`/`relations`: an entity-id ->
+/// 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.
+struct AdjacencyIndex {
+    entity_index: HashMap<u64, usize>,
+    by_entity: HashMap<u64, Vec<usize>>,
+}
+
+impl AdjacencyIndex {
+    fn build(entities: &[Entity], relations: &[Relation]) -> Self {
+        let mut entity_index = HashMap::with_capacity(entities.len());
+        for (i, e) in entities.iter().enumerate() {
+            entity_index.insert(e.id, i);
+        }
+
+        let mut by_entity: HashMap<u64, Vec<usize>> = HashMap::new();
+        for (i, r) in relations.iter().enumerate() {
+            by_entity.entry(r.src).or_default().push(i);
+            if r.tgt != r.src {
+                by_entity.entry(r.tgt).or_default().push(i);
+            }
+        }
+
+        Self {
+            entity_index,
+            by_entity,
+        }
+    }
+
+    /// Indices into `relations` of every edge touching `entity_id`.
+    fn relations_touching(&self, entity_id: u64) -> &[usize] {
+        self.by_entity
+            .get(&entity_id)
+            .map(|v| v.as_slice())
+            .unwrap_or(&[])
+    }
+}
+
 // ---------------------------------------------------------------------------
 // KnowledgeCache
 // ---------------------------------------------------------------------------
@@ -198,6 +251,7 @@ impl KnowledgeCache {
        self.entities.push(Entity {
            id,
            name: name.to_owned(),
+            name_lower: name.to_lowercase(),
            entity_type: entity_type.to_owned(),
            embedding_idx,
            properties: HashMap::new(),
@@ -310,16 +364,22 @@ impl KnowledgeCache {
    ) -> (u64, bool) {
        let lower_name = name.to_lowercase();

-        // Search for the closest existing entity.
-        let best = self
-            .entities
-            .iter()
-            .map(|e| {
-                let dist = levenshtein(&lower_name, &e.name.to_lowercase());
-                (e.id, dist)
-            })
-            .filter(|&(_, dist)| dist <= max_distance)
-            .min_by_key(|&(_, dist)| dist);
+        // Search for the closest existing entity, short-circuiting on an
+        // exact match since no closer candidate can exist.
+        let mut best: Option<(u64, usize)> = None;
+        for e in &self.entities {
+            let dist = levenshtein(&lower_name, &e.name_lower);
+            if dist > max_distance {
+                continue;
+            }
+            if dist == 0 {
+                best = Some((e.id, dist));
+                break;
+            }
+            if best.is_none_or(|(_, best_dist)| dist < best_dist) {
+                best = Some((e.id, dist));
+            }
+        }

        if let Some((id, _)) = best {
            return (id, false);
@@ -337,6 +397,7 @@ 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 mut visited: HashSet<u64> = HashSet::new();
        let mut queue: VecDeque<(u64, usize)> = VecDeque::new();
        let mut results: Vec<(Entity, usize)> = Vec::new();
@@ -349,11 +410,13 @@ impl KnowledgeCache {
                continue;
            }

-            // Collect neighbour IDs from outgoing and incoming edges.
-            let neighbours: Vec<u64> = self
-                .relations
+            // Collect neighbour IDs from outgoing and incoming edges touching
+            // this node only, instead of scanning every relation in the graph.
+            let neighbours: Vec<u64> = idx
+                .relations_touching(current_id)
                .iter()
-                .filter_map(|r| {
+                .filter_map(|&i| {
+                    let r = &self.relations[i];
                    if r.src == current_id {
                        Some(r.tgt)
                    } else if r.tgt == current_id {
@@ -366,9 +429,9 @@ impl KnowledgeCache {

            for neighbour_id in neighbours {
                if visited.insert(neighbour_id)
-                    && let Some(entity) = self.get_entity(neighbour_id)
+                    && let Some(&entity_idx) = idx.entity_index.get(&neighbour_id)
                {
-                    results.push((entity.clone(), depth + 1));
+                    results.push((self.entities[entity_idx].clone(), depth + 1));
                    queue.push_back((neighbour_id, depth + 1));
                }
            }
@@ -439,6 +502,7 @@ impl KnowledgeCache {
        min_activation: f32,
        max_steps: usize,
    ) -> Vec<(u64, f32)> {
+        let idx = AdjacencyIndex::build(&self.entities, &self.relations);
        let mut activation: HashMap<u64, f32> = HashMap::new();

        // Initialise seeds with activation 1.0.
@@ -461,8 +525,10 @@ impl KnowledgeCache {
            let mut any_spread = false;

            for (source_id, source_score) in current {
-                // Spread to all neighbours via outgoing and incoming edges.
-                for rel in &self.relations {
+                // Spread only to edges touching this node, instead of
+                // scanning every relation in the graph per active node.
+                for &rel_idx in idx.relations_touching(source_id) {
+                    let rel = &self.relations[rel_idx];
                    let neighbour_id = if rel.src == source_id {
                        rel.tgt
                    } else if rel.tgt == source_id {
@@ -855,6 +921,19 @@ mod tests {
        assert_eq!(id, orig_id);
    }

+    /// An exact match must win even when a near-match with a smaller Levenshtein
+    /// distance-to-zero gap was scanned first — the early exit on dist == 0
+    /// must not skip past a later exact match.
+    #[test]
+    fn test_resolve_or_create_exact_match_beats_earlier_fuzzy_candidate() {
+        let mut cache = KnowledgeCache::new();
+        cache.add_entity("Alyce", "person", -1); // dist 1 from "Alice"
+        let exact_id = cache.add_entity("Alice", "person", -1); // dist 0
+        let (id, created) = cache.resolve_or_create("Alice", "person", -1, 2);
+        assert!(!created);
+        assert_eq!(id, exact_id);
+    }
+
    #[test]
    fn test_resolve_or_create_no_match_beyond_threshold() {
        let mut cache = KnowledgeCache::new();
@@ -1035,6 +1114,30 @@ mod tests {
        assert!(b_score.unwrap() > 0.0);
    }

+    /// A self-loop relation (src == tgt) must be visited exactly once by the
+    /// adjacency index, matching the pre-index behavior of iterating
+    /// `self.relations` directly (each relation processed once regardless of
+    /// how many of its endpoints match the current node).
+    #[test]
+    fn test_spreading_activation_self_loop_not_double_counted() {
+        let mut cache = KnowledgeCache::new();
+        let a = cache.add_entity("A", "node", -1);
+        cache.add_relation(a, a, "self", 1.0);
+
+        let result = cache.spreading_activation(&[a], 0.5, 0.0001, 1);
+        let a_score = result
+            .iter()
+            .find(|&&(id, _)| id == a)
+            .map(|&(_, s)| s)
+            .unwrap();
+        // Seed activation (1.0) plus exactly one spread contribution
+        // (1.0 * weight 1.0 * decay 0.5), not two.
+        assert!(
+            (a_score - 1.5).abs() < 1e-5,
+            "expected 1.5 (one self-loop contribution), got {a_score}"
+        );
+    }
+
    #[test]
    fn test_spreading_activation_decay_reduces_signal() {
        let mut cache = KnowledgeCache::new();
@@ -466,7 +466,7 @@ 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,
@@ -531,11 +531,14 @@ impl MemoryBackend for ClawhdfBackend {
        query_embedding: &[f32],
        k: usize,
    ) -> Vec<MemorySearchResult> {
-        // 1. Hybrid retrieval (RRF-blended vector + BM25).
+        // 1. Hybrid retrieval (vector + BM25, fused by score).
        let candidates = k.saturating_mul(3).max(10);
-        let raw = self
-            .memory
-            .hybrid_search(query_embedding, query_text, 0.7, 0.3, candidates);
+        let raw = self.memory.hybrid_search_with(
+            query_embedding,
+            query_text,
+            crate::hybrid::DEFAULT_FUSION,
+            candidates,
+        );

        if raw.is_empty() {
            return Vec::new();
@@ -551,6 +554,7 @@ impl MemoryBackend for ClawhdfBackend {
                timestamp: r.timestamp,
                source_channel: r.source_channel.clone(),
                raw_activation: r.activation,
+                relevance: r.score,
            })
            .collect();

@@ -713,11 +717,13 @@ 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)
@@ -748,6 +754,69 @@ impl MemoryBackend for ClawhdfBackend {
    }
 }

+// ─────────────────────────────────────────────────────────────────────────────
+// Ephemeral tier methods on ClawhdfBackend
+// ─────────────────────────────────────────────────────────────────────────────
+
+impl ClawhdfBackend {
+    /// Enable the ephemeral (in-memory only) working memory tier.
+    pub fn enable_ephemeral(&mut self, config: crate::ephemeral::EphemeralConfig) {
+        self.memory.enable_ephemeral(config);
+    }
+
+    /// Store a text value in ephemeral memory.
+    ///
+    /// Returns an error string if the ephemeral tier has not been enabled.
+    pub fn ephemeral_set(
+        &mut self,
+        key: &str,
+        value: &str,
+        ttl_secs: Option<f64>,
+    ) -> Result<(), String> {
+        match self.memory.ephemeral_mut() {
+            Some(s) => {
+                s.set_text(key, value, ttl_secs);
+                Ok(())
+            }
+            None => Err("ephemeral tier not enabled".to_string()),
+        }
+    }
+
+    /// Retrieve a text value from ephemeral memory.
+    ///
+    /// Returns `None` if the tier is disabled, the key is absent, or the
+    /// entry has expired.
+    pub fn ephemeral_get(&mut self, key: &str) -> Option<String> {
+        self.memory
+            .ephemeral_mut()?
+            .get_text(key)
+            .map(|s| s.to_string())
+    }
+
+    /// Delete a key from ephemeral memory.
+    ///
+    /// Returns `true` if the key existed and was removed.
+    pub fn ephemeral_delete(&mut self, key: &str) -> bool {
+        self.memory.ephemeral_mut().is_some_and(|s| s.delete(key))
+    }
+
+    /// Return a snapshot of ephemeral tier statistics, or `None` if the tier
+    /// is not enabled.
+    pub fn ephemeral_stats(&self) -> Option<crate::ephemeral::EphemeralStats> {
+        self.memory.ephemeral().map(|s| s.stats())
+    }
+
+    /// Promote frequently-accessed ephemeral entries to persistent HDF5 storage.
+    ///
+    /// Entries with `access_count >= min_access_count` are moved from the
+    /// ephemeral store into the persistent cache.  Returns the count promoted.
+    pub fn promote_ephemeral(&mut self, min_access_count: u32) -> Result<usize, String> {
+        self.memory
+            .promote_ephemeral(min_access_count)
+            .map_err(|e| e.to_string())
+    }
+}
+
 // ─────────────────────────────────────────────────────────────────────────────
 // Tests
 // ─────────────────────────────────────────────────────────────────────────────
@@ -1333,66 +1402,3 @@ mod tests {
        assert!(out.starts_with("# Title"));
    }
 }
-
-// ─────────────────────────────────────────────────────────────────────────────
-// Ephemeral tier methods on ClawhdfBackend
-// ─────────────────────────────────────────────────────────────────────────────
-
-impl ClawhdfBackend {
-    /// Enable the ephemeral (in-memory only) working memory tier.
-    pub fn enable_ephemeral(&mut self, config: crate::ephemeral::EphemeralConfig) {
-        self.memory.enable_ephemeral(config);
-    }
-
-    /// Store a text value in ephemeral memory.
-    ///
-    /// Returns an error string if the ephemeral tier has not been enabled.
-    pub fn ephemeral_set(
-        &mut self,
-        key: &str,
-        value: &str,
-        ttl_secs: Option<f64>,
-    ) -> Result<(), String> {
-        match self.memory.ephemeral_mut() {
-            Some(s) => {
-                s.set_text(key, value, ttl_secs);
-                Ok(())
-            }
-            None => Err("ephemeral tier not enabled".to_string()),
-        }
-    }
-
-    /// Retrieve a text value from ephemeral memory.
-    ///
-    /// Returns `None` if the tier is disabled, the key is absent, or the
-    /// entry has expired.
-    pub fn ephemeral_get(&mut self, key: &str) -> Option<String> {
-        self.memory
-            .ephemeral_mut()?
-            .get_text(key)
-            .map(|s| s.to_string())
-    }
-
-    /// Delete a key from ephemeral memory.
-    ///
-    /// Returns `true` if the key existed and was removed.
-    pub fn ephemeral_delete(&mut self, key: &str) -> bool {
-        self.memory.ephemeral_mut().is_some_and(|s| s.delete(key))
-    }
-
-    /// Return a snapshot of ephemeral tier statistics, or `None` if the tier
-    /// is not enabled.
-    pub fn ephemeral_stats(&self) -> Option<crate::ephemeral::EphemeralStats> {
-        self.memory.ephemeral().map(|s| s.stats())
-    }
-
-    /// Promote frequently-accessed ephemeral entries to persistent HDF5 storage.
-    ///
-    /// Entries with `access_count >= min_access_count` are moved from the
-    /// ephemeral store into the persistent cache.  Returns the count promoted.
-    pub fn promote_ephemeral(&mut self, min_access_count: u32) -> Result<usize, String> {
-        self.memory
-            .promote_ephemeral(min_access_count)
-            .map_err(|e| e.to_string())
-    }
-}
@@ -105,6 +105,23 @@ impl ProvenanceStore {
        self.records.insert(provenance.record_id, provenance);
    }

+    /// Renumber records after the store was compacted. `index_map[old]` is
+    /// the record's new id, or `None` if it was removed. Without this, every
+    /// surviving record's hash ends up filed under some other record's id and
+    /// the next integrity check reports a bogus mismatch.
+    pub fn remap(&mut self, index_map: &[Option<usize>]) {
+        let old = std::mem::take(&mut self.records);
+        for (old_id, mut prov) in old {
+            let new_id = usize::try_from(old_id)
+                .ok()
+                .and_then(|i| index_map.get(i).copied().flatten());
+            if let Some(new_id) = new_id {
+                prov.record_id = new_id as u64;
+                self.records.insert(new_id as u64, prov);
+            }
+        }
+    }
+
    /// Retrieve by record ID.
    pub fn get(&self, record_id: u64) -> Option<&MemoryProvenance> {
        self.records.get(&record_id)
@@ -6,6 +6,11 @@
 //! - Temporal expansion (time-related rewrites)
 //! - Morphological variants (stemming-like transforms)
 //! - Knowledge graph expansion (entity aliases and neighbors)
+//!
+//! The morphological rules are crude suffix swaps, so some variants are not
+//! words ("during" -> "dured"). That is tolerable for a BM25 stage, which
+//! simply finds no postings for a nonsense term, but it means expansion is not
+//! free: measure before enabling it on a retrieval path.

 use crate::knowledge::KnowledgeCache;

@@ -340,20 +345,87 @@ fn contains_phrase(text: &str, phrase: &str) -> bool {

 /// Replace a phrase in `text` case-insensitively, preserving surrounding case.
 fn replace_word_case_insensitive(text: &str, from: &str, to: &str) -> String {
-    case_insensitive_replace(text, from, to)
+    replace_first(text, from, to, MatchKind::WholeWord)
 }

 fn case_insensitive_replace(text: &str, from: &str, to: &str) -> String {
-    let lower = text.to_lowercase();
-    let lower_from = from.to_lowercase();
-    if let Some(pos) = lower.find(&lower_from) {
-        let end = pos + from.len();
-        format!("{}{}{}", &text[..pos], to, &text[end..])
-    } else {
-        text.to_string()
+    replace_first(text, from, to, MatchKind::Substring)
+}
+
+/// Whether a match may fall inside a larger word.
+#[derive(Clone, Copy, PartialEq)]
+enum MatchKind {
+    /// Match anywhere, including inside another word.
+    Substring,
+    /// Match only when both ends sit on a word boundary.
+    WholeWord,
+}
+
+/// Replace the first case-insensitive match of `from` in `text` with `to`.
+///
+/// Matching walks the *original* string rather than a lowercased copy. The
+/// previous implementation searched `text.to_lowercase()` and then sliced
+/// `text` with the offsets it found, which only holds while lowercasing
+/// preserves byte length. It does not: Turkish `İ` (2 bytes) lowercases to
+/// `i` + U+0307 (3 bytes), so every later offset was wrong — silently
+/// corrupting the output, or panicking when an offset landed inside a
+/// character or past the end. `"İ AI"` was enough to panic.
+fn replace_first(text: &str, from: &str, to: &str, kind: MatchKind) -> String {
+    match find_case_insensitive(text, from, kind) {
+        Some((start, end)) => {
+            let mut out = String::with_capacity(text.len() - (end - start) + to.len());
+            out.push_str(&text[..start]);
+            out.push_str(to);
+            out.push_str(&text[end..]);
+            out
+        }
+        None => text.to_string(),
    }
 }

+/// Byte range of the first case-insensitive match of `needle` in `haystack`.
+fn find_case_insensitive(haystack: &str, needle: &str, kind: MatchKind) -> Option<(usize, usize)> {
+    if needle.is_empty() {
+        return None;
+    }
+    let lowered: Vec<char> = needle.chars().flat_map(char::to_lowercase).collect();
+    let is_word = |c: char| c.is_alphanumeric() || c == '_';
+
+    for (start, _) in haystack.char_indices() {
+        if kind == MatchKind::WholeWord
+            && haystack[..start].chars().next_back().is_some_and(is_word)
+        {
+            continue; // mid-word: "ai" inside "training"
+        }
+        let mut matched = 0usize;
+        let mut end = start;
+        for (offset, ch) in haystack[start..].char_indices() {
+            if matched == lowered.len() {
+                break;
+            }
+            let mut consumed_all = true;
+            for lc in ch.to_lowercase() {
+                if lowered.get(matched) != Some(&lc) {
+                    consumed_all = false;
+                    break;
+                }
+                matched += 1;
+            }
+            if !consumed_all {
+                break;
+            }
+            end = start + offset + ch.len_utf8();
+        }
+        if matched == lowered.len()
+            && !(kind == MatchKind::WholeWord
+                && haystack[end..].chars().next().is_some_and(is_word))
+        {
+            return Some((start, end));
+        }
+    }
+    None
+}
+
 /// Simple whitespace/punctuation tokenizer.
 fn tokenize(text: &str) -> Vec<String> {
    text.split(|c: char| !c.is_alphanumeric())
@@ -637,4 +709,86 @@ mod tests {
            expanded.iter().map(|x| &x.text).collect::<Vec<_>>()
        );
    }
+    #[test]
+    fn acronyms_only_match_whole_words() {
+        let ex = QueryExpander::new(QueryExpansionConfig::default());
+        // "training" contains "ai", "programming" contains "pr". These used to
+        // be rewritten to "trArtificial Intelligencening" and
+        // "Pull Requestogramming".
+        for query in [
+            "How many miles during my marathon training?",
+            "Which programming language did I pick?",
+            "I updated the maintainer list",
+        ] {
+            for expansion in ex.expand(query) {
+                assert!(
+                    expansion.expansion_type != "acronym",
+                    "{query:?} produced {expansion:?}"
+                );
+            }
+        }
+        // A real acronym still expands, in both directions.
+        let texts: Vec<String> = ex
+            .expand("What about the API and the database?")
+            .into_iter()
+            .filter(|e| e.expansion_type == "acronym")
+            .map(|e| e.text)
+            .collect();
+        assert!(
+            texts
+                .iter()
+                .any(|t| t.contains("Application Programming Interface")),
+            "{texts:?}"
+        );
+        assert!(texts.iter().any(|t| t.contains("DB")), "{texts:?}");
+    }
+
+    #[test]
+    fn non_ascii_queries_do_not_panic_or_corrupt() {
+        let ex = QueryExpander::new(QueryExpansionConfig::default());
+        // Turkish 'İ' is 2 bytes but lowercases to 3, so offsets taken from a
+        // lowercased copy no longer line up with the original. `"İ AI"` used
+        // to panic; `"İstanbul AI trip"` used to silently eat a character.
+        for query in ["İ AI", "İé AI", "İİ ML", "İstanbul AI trip", "ǰ ML notes"] {
+            for expansion in ex.expand(query) {
+                assert!(
+                    expansion.text.contains('İ') || expansion.text.contains('ǰ'),
+                    "{query:?} lost its leading character: {expansion:?}"
+                );
+            }
+        }
+        let expanded = ex.expand("İstanbul AI trip");
+        assert!(
+            expanded
+                .iter()
+                .any(|e| e.text == "İstanbul Artificial Intelligence trip"),
+            "{expanded:?}"
+        );
+    }
+
+    #[test]
+    fn whole_word_matching_handles_string_edges_and_case() {
+        assert_eq!(
+            replace_word_case_insensitive("ai tools", "AI", "Artificial Intelligence"),
+            "Artificial Intelligence tools"
+        );
+        assert_eq!(
+            replace_word_case_insensitive("tools for ai", "AI", "Artificial Intelligence"),
+            "tools for Artificial Intelligence"
+        );
+        assert_eq!(
+            replace_word_case_insensitive("the aim", "AI", "Artificial Intelligence"),
+            "the aim",
+            "must not match inside a word"
+        );
+        assert_eq!(
+            replace_word_case_insensitive("no match here", "xyz", "abc"),
+            "no match here"
+        );
+        // Only the first occurrence is replaced, as before.
+        assert_eq!(
+            replace_word_case_insensitive("ai and ai", "ai", "ML"),
+            "ML and ai"
+        );
+    }
 }
@@ -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);
@@ -12,10 +12,18 @@ use crate::MemoryError;
 use crate::cache::MemoryCache;
 use crate::knowledge::KnowledgeCache;
 use crate::session::SessionCache;
+use crate::wal::WalMark;

 pub const SCHEMA_VERSION: &str = "1.0";
 pub const ZEROCLAW_VERSION: &str = "0.8.0";

+/// `/meta` attributes holding the [`WalMark`] of the WAL prefix already folded
+/// into this file. Absent on files written before the mark existed, and when
+/// the checkpoint was taken with an empty WAL.
+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";
+
 /// Build a complete HDF5 file from the in-memory state.
 pub fn build_hdf5_file(
    config: &MemoryConfig,
@@ -23,6 +31,47 @@ pub fn build_hdf5_file(
    sessions: &SessionCache,
    knowledge: &KnowledgeCache,
 ) -> Result<Vec<u8>, MemoryError> {
+    build_hdf5_file_with_mark(config, cache, sessions, knowledge, None)
+}
+
+/// [`build_hdf5_file`], recording which WAL prefix this state already
+/// contains (see [`WalMark`]) so a crash before the WAL is truncated doesn't
+/// replay those entries a second time.
+pub fn build_hdf5_file_with_mark(
+    config: &MemoryConfig,
+    cache: &MemoryCache,
+    sessions: &SessionCache,
+    knowledge: &KnowledgeCache,
+    wal_applied: Option<WalMark>,
+) -> Result<Vec<u8>, MemoryError> {
+    let meta = CheckpointMeta {
+        wal_applied,
+        ann_generation: None,
+    };
+    build_hdf5_file_with_meta(config, cache, sessions, knowledge, &meta)
+}
+
+/// Bookkeeping a checkpoint records in `/meta` beside the store's contents.
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
+pub struct CheckpointMeta {
+    /// The WAL prefix this checkpoint already contains; see [`WalMark`].
+    pub wal_applied: Option<WalMark>,
+    /// Identifies the vector-index sidecar (`<store>.h5.ann`) written with this
+    /// checkpoint. A sidecar is loaded only if it carries the same value, so
+    /// one left over from another checkpoint can never be attached to records
+    /// it wasn't built from.
+    pub ann_generation: Option<u64>,
+}
+
+/// [`build_hdf5_file`] with checkpoint bookkeeping.
+pub fn build_hdf5_file_with_meta(
+    config: &MemoryConfig,
+    cache: &MemoryCache,
+    sessions: &SessionCache,
+    knowledge: &KnowledgeCache,
+    checkpoint: &CheckpointMeta,
+) -> Result<Vec<u8>, MemoryError> {
+    let wal_applied = checkpoint.wal_applied;
    let mut builder = clawhdf5::FileBuilder::new();

    // /meta group with schema attributes
@@ -34,10 +83,44 @@ pub fn build_hdf5_file(
    meta.set_attr("embedding_dim", AttrValue::I64(config.embedding_dim as i64));
    meta.set_attr("chunk_size", AttrValue::I64(config.chunk_size as i64));
    meta.set_attr("overlap", AttrValue::I64(config.overlap as i64));
+    // Behavioural settings. These used to live only in memory, so reopening a
+    // store silently reset them to defaults — e.g. a compressed store was
+    // rewritten uncompressed by the first checkpoint after a reopen. Loaders
+    // treat each one as optional so older files keep opening.
+    meta.set_attr("float16", AttrValue::I64(config.float16.into()));
+    meta.set_attr("compression", AttrValue::I64(config.compression.into()));
+    meta.set_attr(
+        "compression_level",
+        AttrValue::I64(config.compression_level.into()),
+    );
+    meta.set_attr(
+        "compact_threshold",
+        AttrValue::F64(config.compact_threshold.into()),
+    );
+    meta.set_attr("hebbian_boost", AttrValue::F64(config.hebbian_boost.into()));
+    meta.set_attr("decay_factor", AttrValue::F64(config.decay_factor.into()));
+    meta.set_attr("wal_enabled", AttrValue::I64(config.wal_enabled.into()));
+    meta.set_attr(
+        "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(
        "edgehdf5_version",
        AttrValue::String(ZEROCLAW_VERSION.into()),
    );
+    if let Some(mark) = wal_applied.filter(|m| m.len > 0) {
+        meta.set_attr(WAL_APPLIED_LEN_ATTR, AttrValue::I64(mark.len as i64));
+        meta.set_attr(WAL_APPLIED_CRC_ATTR, AttrValue::I64(i64::from(mark.crc)));
+    }
+    if let Some(generation) = checkpoint.ann_generation {
+        // Stored as the i64 with the same bits; attributes have no u64 scalar
+        // round trip through every reader.
+        meta.set_attr(ANN_GENERATION_ATTR, AttrValue::I64(generation as i64));
+    }
    // 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();
@@ -74,7 +157,7 @@ fn build_memory_group(
    {
        let ds = group
            .create_dataset("embeddings")
-            .with_f32_data(&flat)
+            .with_f32_data(flat)
            .with_shape(&[n, d]);

        // Chunk size tuning: target ~256KB per chunk for optimal I/O
@@ -83,15 +166,33 @@ fn build_memory_group(
            let rows_per_chunk = (target_chunk_bytes / (d * 4)).max(1).min(n);
            ds.with_chunks(&[rows_per_chunk, d]);

-            // Compression: Zstd for embeddings — faster than deflate at same ratio.
-            // Shuffle is applied automatically (auto-shuffle pre-filter).
+            // Compression. Shuffle is applied automatically (auto-shuffle
+            // pre-filter). Zstd is faster than deflate at the same ratio but
+            // pulls in libzstd, so it is opt-in via the `zstd` feature; the
+            // default build uses deflate, which is always available. (This
+            // used to call `with_zstd` unconditionally, so without the
+            // feature every checkpoint of a compressed store failed with
+            // "unsupported filter: 32015".) Both are standard HDF5 filters;
+            // reading a zstd-compressed store needs a zstd-enabled build.
            if config.compression {
-                let level = if config.compression_level > 0 {
-                    config.compression_level.min(22)
-                } else {
-                    3 // Zstd level 3: fast + good ratio for f32 embeddings
-                };
-                ds.with_zstd(level);
+                #[cfg(feature = "zstd")]
+                {
+                    let level = if config.compression_level > 0 {
+                        config.compression_level.min(22)
+                    } else {
+                        3 // fast + good ratio for f32 embeddings
+                    };
+                    ds.with_zstd(level);
+                }
+                #[cfg(not(feature = "zstd"))]
+                {
+                    let level = if config.compression_level > 0 {
+                        config.compression_level.min(9)
+                    } else {
+                        4
+                    };
+                    ds.with_deflate(level);
+                }
            }
        }

@@ -309,6 +410,36 @@ fn write_string_dataset(
 }

 /// 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 len = match attrs.get(WAL_APPLIED_LEN_ATTR)? {
+        AttrValue::I64(v) => u64::try_from(*v).ok()?,
+        _ => return None,
+    };
+    let crc = match attrs.get(WAL_APPLIED_CRC_ATTR)? {
+        AttrValue::I64(v) => u32::try_from(*v).ok()?,
+        _ => return None,
+    };
+    Some(WalMark { len, crc })
+}
+
+/// Read the checkpoint bookkeeping from `/meta`.
+pub fn read_checkpoint_meta(file: &clawhdf5::File) -> CheckpointMeta {
+    let ann_generation = file
+        .group("meta")
+        .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,
+        });
+    CheckpointMeta {
+        wal_applied: read_wal_mark(file),
+        ann_generation,
+    }
+}
+
 pub fn validate_and_load(
    file: &clawhdf5::File,
 ) -> Result<(MemoryConfig, MemoryCache, SessionCache, KnowledgeCache), MemoryError> {
@@ -344,15 +475,20 @@ pub fn validate_and_load(
        embedding_dim,
        chunk_size,
        overlap,
-        float16: false,
-        compression: false,
-        compression_level: 0,
-        compact_threshold: 0.3,
-        hebbian_boost: 0.15,
-        decay_factor: 0.98,
+        float16: optional_bool_attr(&attrs, "float16", false),
+        compression: optional_bool_attr(&attrs, "compression", false),
+        compression_level: optional_i64_attr(&attrs, "compression_level")
+            .and_then(|v| u32::try_from(v).ok())
+            .unwrap_or(0),
+        compact_threshold: optional_f32_attr(&attrs, "compact_threshold", 0.3),
+        hebbian_boost: optional_f32_attr(&attrs, "hebbian_boost", 0.15),
+        decay_factor: optional_f32_attr(&attrs, "decay_factor", 0.98),
        created_at,
-        wal_enabled: true,
-        wal_max_entries: 500,
+        wal_enabled: optional_bool_attr(&attrs, "wal_enabled", true),
+        wal_max_entries: optional_i64_attr(&attrs, "wal_max_entries")
+            .and_then(|v| usize::try_from(v).ok())
+            .unwrap_or(500),
+        quantized_index: optional_bool_attr(&attrs, "quantized_index", false),
    };

    // Load /memory group
@@ -391,27 +527,48 @@ fn load_memory_group(
    let tags = read_string_dataset_from_group(&group, "tags")?;
    let tombstones = read_u8_dataset(&group, "tombstones")?;

-    // Read norms if present, otherwise compute from embeddings
-    let norms = match read_f32_dataset(&group, "norms") {
-        Ok(n) if n.len() == n.len() => n,
-        _ => {
-            // Compute norms from flat embeddings
-            flat_embeddings
-                .chunks(embedding_dim)
-                .map(|chunk| {
-                    let sq_sum: f32 = chunk.iter().map(|x| x * x).sum();
-                    sq_sum.sqrt()
-                })
-                .collect()
+    // Every per-record dataset must describe exactly `n` records. Without
+    // this, a truncated or hand-edited file loads "successfully" and then
+    // panics on the first out-of-bounds index during search/delete.
+    if embedding_dim == 0 {
+        return Err(MemoryError::Schema(format!(
+            "/memory has {n} records but embedding_dim is 0"
+        )));
+    }
+    let expected_flat = n.checked_mul(embedding_dim).ok_or_else(|| {
+        MemoryError::Schema(format!("/memory size overflow: {n} x {embedding_dim}"))
+    })?;
+    let check_len = |name: &str, actual: usize, expected: usize| {
+        if actual == expected {
+            Ok(())
+        } else {
+            Err(MemoryError::Schema(format!(
+                "/memory/{name} has {actual} entries, expected {expected} \
+                 ({n} records)"
+            )))
        }
    };
+    check_len("embeddings", flat_embeddings.len(), expected_flat)?;
+    check_len("source_channel", source_channels.len(), n)?;
+    check_len("timestamps", timestamps.len(), n)?;
+    check_len("session_ids", session_ids.len(), n)?;
+    check_len("tags", tags.len(), n)?;
+    check_len("tombstones", tombstones.len(), n)?;

-    // Unflatten embeddings
-    let embeddings: Vec<Vec<f32>> = flat_embeddings
-        .chunks(embedding_dim)
-        .map(|c| c.to_vec())
-        .collect();
+    // Norms are derived data: use the stored ones only if they are present
+    // and the right length, otherwise recompute from the embeddings.
+    let norms = match read_f32_dataset(&group, "norms") {
+        Ok(stored) if stored.len() == n => stored,
+        _ => flat_embeddings
+            .chunks(embedding_dim)
+            .map(|chunk| {
+                let sq_sum: f32 = chunk.iter().map(|x| x * x).sum();
+                sq_sum.sqrt()
+            })
+            .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,
@@ -419,7 +576,7 @@ 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;
@@ -480,6 +637,7 @@ fn load_knowledge_group(file: &clawhdf5::File) -> Result<KnowledgeCache, MemoryE
            cache.entities.push(crate::knowledge::Entity {
                id: entity_ids[i] as u64,
                name: entity_names[i].clone(),
+                name_lower: entity_names[i].to_lowercase(),
                entity_type: entity_types[i].clone(),
                embedding_idx: emb_idxs[i],
                ..Default::default()
@@ -529,6 +687,27 @@ fn extract_string_attr(
    }
 }

+type MetaAttrs = std::collections::HashMap<String, AttrValue>;
+
+fn optional_i64_attr(attrs: &MetaAttrs, name: &str) -> Option<i64> {
+    match attrs.get(name) {
+        Some(AttrValue::I64(v)) => Some(*v),
+        _ => None,
+    }
+}
+
+fn optional_bool_attr(attrs: &MetaAttrs, name: &str, default: bool) -> bool {
+    optional_i64_attr(attrs, name).map_or(default, |v| v != 0)
+}
+
+/// Finite values only: a NaN threshold/decay would poison every comparison.
+fn optional_f32_attr(attrs: &MetaAttrs, name: &str, default: f32) -> f32 {
+    match attrs.get(name) {
+        Some(AttrValue::F64(v)) if v.is_finite() => *v as f32,
+        _ => default,
+    }
+}
+
 fn extract_i64_attr(
    attrs: &std::collections::HashMap<String, AttrValue>,
    name: &str,
@@ -614,3 +793,108 @@ fn read_u8_dataset(group: &clawhdf5::Group<'_>, name: &str) -> Result<Vec<u8>, M
        .map_err(|e| MemoryError::Hdf5(format!("cannot read u8 from {name}: {e}")))?;
    Ok(data.into_iter().map(|v| v as u8).collect())
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn config() -> MemoryConfig {
+        MemoryConfig::new(std::path::PathBuf::from("unused.h5"), "agent", 4)
+    }
+
+    fn cache_with(n: usize) -> MemoryCache {
+        let mut cache = MemoryCache::new(4);
+        for i in 0..n {
+            cache.push(
+                format!("chunk {i}"),
+                vec![i as f32 + 1.0, 0.0, 0.0, 0.0],
+                "user".into(),
+                i as f64,
+                "s".into(),
+                "t".into(),
+            );
+        }
+        cache
+    }
+
+    fn roundtrip(cache: &MemoryCache) -> Result<MemoryCache, MemoryError> {
+        let bytes = build_hdf5_file(
+            &config(),
+            cache,
+            &SessionCache::new(),
+            &KnowledgeCache::new(),
+        )?;
+        let file =
+            clawhdf5::File::from_bytes(bytes).map_err(|e| MemoryError::Hdf5(e.to_string()))?;
+        validate_and_load(&file).map(|(_, cache, _, _)| cache)
+    }
+
+    #[test]
+    fn behavioural_config_survives_a_reopen() {
+        let mut cfg = config();
+        cfg.compression = true;
+        cfg.compression_level = 7;
+        cfg.compact_threshold = 0.5;
+        cfg.hebbian_boost = 0.25;
+        cfg.decay_factor = 0.9;
+        cfg.wal_enabled = false;
+        cfg.wal_max_entries = 42;
+        let bytes = build_hdf5_file(
+            &cfg,
+            &cache_with(2),
+            &SessionCache::new(),
+            &KnowledgeCache::new(),
+        )
+        .unwrap();
+        let file = clawhdf5::File::from_bytes(bytes).unwrap();
+        let (loaded, loaded_cache, ..) = validate_and_load(&file).unwrap();
+        // The compressed embeddings must also read back intact.
+        assert_eq!(loaded_cache.embeddings, cache_with(2).embeddings);
+        assert!(loaded.compression);
+        assert_eq!(loaded.compression_level, 7);
+        assert_eq!(loaded.compact_threshold, 0.5);
+        assert_eq!(loaded.hebbian_boost, 0.25);
+        assert_eq!(loaded.decay_factor, 0.9);
+        assert!(!loaded.wal_enabled);
+        assert_eq!(loaded.wal_max_entries, 42);
+    }
+
+    #[test]
+    fn consistent_store_loads() {
+        let loaded = roundtrip(&cache_with(3)).unwrap();
+        assert_eq!(loaded.chunks.len(), 3);
+        assert_eq!(loaded.norms, vec![1.0, 2.0, 3.0]);
+    }
+
+    #[test]
+    fn wrong_length_norms_are_recomputed_not_trusted() {
+        // Regression: the guard used to be `n.len() == n.len()`, so a norms
+        // dataset of any length was accepted and corrupted every cosine score.
+        let mut cache = cache_with(3);
+        cache.norms = vec![99.0];
+        let loaded = roundtrip(&cache).unwrap();
+        assert_eq!(loaded.norms, vec![1.0, 2.0, 3.0]);
+    }
+
+    #[test]
+    fn mismatched_per_record_datasets_are_schema_errors() {
+        type Corrupt = fn(&mut MemoryCache);
+        let cases: [(&str, Corrupt); 5] = [
+            ("tombstones", |c| c.tombstones.truncate(1)),
+            ("timestamps", |c| c.timestamps.truncate(1)),
+            ("tags", |c| c.tags.truncate(1)),
+            ("session_ids", |c| c.session_ids.truncate(1)),
+            ("source_channel", |c| c.source_channels.truncate(1)),
+        ];
+        for (name, corrupt) in cases {
+            let mut cache = cache_with(3);
+            corrupt(&mut cache);
+            match roundtrip(&cache) {
+                Err(MemoryError::Schema(msg)) => {
+                    assert!(msg.contains(name), "{name}: unexpected message {msg}")
+                }
+                other => panic!("{name}: expected Schema error, got {:?}", other.map(|_| ())),
+            }
+        }
+    }
+}
@@ -4,7 +4,7 @@ use std::path::Path;

 use crate::bm25;
 use crate::hybrid;
-use crate::{HDF5Memory, MemoryError, Result, SearchResult};
+use crate::{HDF5Memory, MAX_ACTIVATION_WEIGHT, MemoryError, Result, SearchResult};

 impl HDF5Memory {
    /// Vector + keyword scoring stage of [`HDF5Memory::hybrid_search`].
@@ -20,8 +20,7 @@ impl HDF5Memory {
        query_embedding: &[f32],
        query_text: &str,
        bm25: &bm25::BM25Index,
-        vector_weight: f32,
-        keyword_weight: f32,
+        fusion: hybrid::Fusion,
        k: usize,
    ) -> Vec<(usize, f32)> {
        self.ensure_hnsw_fresh();
@@ -30,29 +29,40 @@ impl HDF5Memory {
                // 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 candidates = index.search(query_embedding, pool, pool);
+                // 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))
+                    .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();
-                let kw_scores = bm25.search(query_text, self.cache.len());
-                hybrid::merge_vector_keyword(
-                    vec_scores,
-                    kw_scores,
-                    vector_weight,
-                    keyword_weight,
-                    k,
-                )
+                // Fusion normalises over every keyword match, so it needs all
+                // the scores — but not ranked.
+                let kw_scores = bm25.scores(query_text);
+                hybrid::fuse(vec_scores, kw_scores, fusion, k)
            }
-            _ => hybrid::hybrid_search(
+            _ => hybrid::hybrid_search_fused(
                query_embedding,
                query_text,
                &self.cache.embeddings,
                &self.cache.chunks,
                &self.cache.tombstones,
                bm25,
-                vector_weight,
-                keyword_weight,
+                fusion,
                k,
            ),
        }
@@ -64,19 +74,17 @@ impl HDF5Memory {
        query_embedding: &[f32],
        query_text: &str,
        bm25: &bm25::BM25Index,
-        vector_weight: f32,
-        keyword_weight: f32,
+        fusion: hybrid::Fusion,
        k: usize,
    ) -> Vec<(usize, f32)> {
-        hybrid::hybrid_search(
+        hybrid::hybrid_search_fused(
            query_embedding,
            query_text,
            &self.cache.embeddings,
            &self.cache.chunks,
            &self.cache.tombstones,
            bm25,
-            vector_weight,
-            keyword_weight,
+            fusion,
            k,
        )
    }
@@ -90,15 +98,35 @@ impl HDF5Memory {
        keyword_weight: f32,
        k: usize,
    ) -> Vec<SearchResult> {
-        let bm25 = bm25::BM25Index::build(&self.cache.chunks, &self.cache.tombstones);
-        let scored = self.vector_keyword_search(
+        self.hybrid_search_with(
            query_embedding,
            query_text,
-            &bm25,
-            vector_weight,
-            keyword_weight,
+            hybrid::Fusion::Weighted {
+                vector: vector_weight,
+                keyword: keyword_weight,
+            },
            k,
-        );
+        )
+    }
+
+    /// [`HDF5Memory::hybrid_search`] with the fusion method chosen explicitly.
+    ///
+    /// [`hybrid::DEFAULT_FUSION`] is what the weighted form defaults to;
+    /// [`hybrid::Fusion::Rrf`] combines the two stages by rank instead of by
+    /// score.
+    pub fn hybrid_search_with(
+        &mut self,
+        query_embedding: &[f32],
+        query_text: &str,
+        fusion: hybrid::Fusion,
+        k: usize,
+    ) -> Vec<SearchResult> {
+        // 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 mut results: Vec<SearchResult> = scored
            .into_iter()
            .map(|(idx, score)| {
@@ -113,23 +141,45 @@ impl HDF5Memory {
                }
            })
            .collect();
+        // Ties broken by index so results (and therefore which records get
+        // boosted) don't depend on HashMap iteration order upstream.
        results.sort_by(|a, b| {
            b.score
                .partial_cmp(&a.score)
                .unwrap_or(std::cmp::Ordering::Equal)
+                .then(a.index.cmp(&b.index))
        });

-        let hit_indices: Vec<usize> = results.iter().map(|r| r.index).collect();
+        // 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
+        // important just because they were nearby in iteration order.
+        let hit_indices: Vec<usize> = results
+            .iter()
+            .filter(|r| r.score > 0.0)
+            .map(|r| r.index)
+            .collect();
        self.apply_hebbian_boost(&hit_indices);
-        self.flush().ok();
+        self.bm25 = Some(bm25);

        results
    }

+    /// 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
+    /// `hybrid_search` cost O(store size) in disk I/O. They are a ranking hint,
+    /// not user data: a crash before the next checkpoint only forgets the
+    /// boosts since the last one.
    fn apply_hebbian_boost(&mut self, hit_indices: &[usize]) {
-        for &idx in hit_indices {
-            self.cache.activation_weights[idx] += self.config.hebbian_boost;
+        if hit_indices.is_empty() || self.config.hebbian_boost == 0.0 {
+            return;
        }
+        for &idx in hit_indices {
+            let w = &mut self.cache.activation_weights[idx];
+            *w = (*w + self.config.hebbian_boost).min(MAX_ACTIVATION_WEIGHT);
+        }
+        self.activations_dirty = true;
    }

    /// Get the chunk text for a memory entry by index.
@@ -11,6 +11,7 @@ use crate::cache::MemoryCache;
 use crate::knowledge::KnowledgeCache;
 use crate::schema;
 use crate::session::SessionCache;
+use crate::wal::WalMark;

 /// Write all in-memory state to an HDF5 file on disk.
 pub fn write_to_disk(
@@ -20,7 +21,36 @@ pub fn write_to_disk(
    sessions: &SessionCache,
    knowledge: &KnowledgeCache,
 ) -> Result<(), MemoryError> {
-    let bytes = schema::build_hdf5_file(config, cache, sessions, knowledge)?;
+    write_to_disk_with_mark(path, config, cache, sessions, knowledge, None)
+}
+
+/// [`write_to_disk`] for a checkpoint: `wal_applied` is the mark of the WAL
+/// prefix whose entries `cache` already contains.
+pub fn write_to_disk_with_mark(
+    path: &Path,
+    config: &MemoryConfig,
+    cache: &MemoryCache,
+    sessions: &SessionCache,
+    knowledge: &KnowledgeCache,
+    wal_applied: Option<WalMark>,
+) -> Result<(), MemoryError> {
+    let meta = schema::CheckpointMeta {
+        wal_applied,
+        ann_generation: None,
+    };
+    write_to_disk_with_meta(path, config, cache, sessions, knowledge, &meta)
+}
+
+/// [`write_to_disk`] with full checkpoint bookkeeping.
+pub fn write_to_disk_with_meta(
+    path: &Path,
+    config: &MemoryConfig,
+    cache: &MemoryCache,
+    sessions: &SessionCache,
+    knowledge: &KnowledgeCache,
+    checkpoint: &schema::CheckpointMeta,
+) -> Result<(), MemoryError> {
+    let bytes = schema::build_hdf5_file_with_meta(config, cache, sessions, knowledge, checkpoint)?;

    if bytes.is_empty() {
        return Err(MemoryError::Hdf5("build_hdf5_file produced 0 bytes".into()));
@@ -28,9 +58,41 @@ pub fn write_to_disk(

    // Write to a temp file first, then rename for atomicity
    let tmp_path = path.with_extension("h5.tmp");
-    std::fs::write(&tmp_path, &bytes).map_err(MemoryError::Io)?;
-    std::fs::rename(&tmp_path, path).map_err(MemoryError::Io)?;
+    write_synced(&tmp_path, &bytes)?;
+    rename_synced(&tmp_path, path)
+}

+/// Write `bytes` to `path` and flush them to stable storage.
+pub(crate) fn write_synced(path: &Path, bytes: &[u8]) -> Result<(), MemoryError> {
+    use std::io::Write;
+    let mut f = std::fs::File::create(path).map_err(MemoryError::Io)?;
+    f.write_all(bytes).map_err(MemoryError::Io)?;
+    f.sync_all().map_err(MemoryError::Io)
+}
+
+/// Rename `from` over `to`, then sync the parent directory so the rename
+/// itself survives a power loss. `from` must already be synced: without that,
+/// the rename can reach disk before the data and leave an empty or partial
+/// file under the final name.
+///
+/// This is per-checkpoint/snapshot cost only (each is already a full file
+/// write). Individual WAL appends are deliberately not synced — see the
+/// durability notes in the crate docs.
+pub(crate) fn rename_synced(from: &Path, to: &Path) -> Result<(), MemoryError> {
+    std::fs::rename(from, to).map_err(MemoryError::Io)?;
+    #[cfg(unix)]
+    if let Some(dir) = to.parent() {
+        let dir = if dir.as_os_str().is_empty() {
+            Path::new(".")
+        } else {
+            dir
+        };
+        // Directory fsync is best-effort: some filesystems refuse it, and the
+        // rename has already happened.
+        if let Ok(d) = std::fs::File::open(dir) {
+            let _ = d.sync_all();
+        }
+    }
    Ok(())
 }

@@ -42,6 +104,15 @@ pub fn write_to_disk(
 pub fn read_from_disk(
    path: &Path,
 ) -> Result<(MemoryConfig, MemoryCache, SessionCache, KnowledgeCache), MemoryError> {
+    read_from_disk_with_mark(path).map(|(state, _mark)| state)
+}
+
+/// Everything [`read_from_disk`] returns.
+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
@@ -53,8 +124,23 @@ pub fn read_from_disk(

    let (mut config, cache, sessions, knowledge) = schema::validate_and_load(&file)?;
    config.path = path.to_path_buf();
+    let wal_applied = schema::read_wal_mark(&file);

-    Ok((config, cache, sessions, knowledge))
+    Ok(((config, cache, sessions, knowledge), wal_applied))
+}
+
+/// [`read_from_disk`], plus all checkpoint bookkeeping.
+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())
+        .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();
+    let meta = schema::read_checkpoint_meta(&file);
+    Ok(((config, cache, sessions, knowledge), meta))
 }

 /// Copy an HDF5 file atomically to a destination.
@@ -78,7 +164,10 @@ pub fn snapshot_file(src: &Path, dest: &Path) -> Result<std::path::PathBuf, Memo
    // Atomic copy: write to temp, then rename
    let tmp_path = dest_file.with_extension("h5.tmp");
    std::fs::copy(src, &tmp_path).map_err(MemoryError::Io)?;
-    std::fs::rename(&tmp_path, &dest_file).map_err(MemoryError::Io)?;
+    std::fs::File::open(&tmp_path)
+        .and_then(|f| f.sync_all())
+        .map_err(MemoryError::Io)?;
+    rename_synced(&tmp_path, &dest_file)?;

    Ok(dest_file)
 }
@@ -0,0 +1,79 @@
+//! Single-writer guard for a memory store.
+//!
+//! `HDF5Memory` keeps the whole store in memory and rewrites the `.h5` file at
+//! every checkpoint, so two handles on one store (two processes, or two opens
+//! in one process) silently destroy each other's data: whoever checkpoints
+//! last wins, and both append to the same WAL with independent CRC chains.
+//! The lock turns that into an immediate, explicit error.
+
+use std::fs::{File, OpenOptions, TryLockError};
+use std::path::{Path, PathBuf};
+
+use crate::MemoryError;
+
+const LOCK_RETRIES: u32 = 25;
+const LOCK_RETRY_DELAY: std::time::Duration = std::time::Duration::from_millis(10);
+
+/// An exclusive advisory lock on `<store>.h5.lock`, held for the lifetime of
+/// the owning `HDF5Memory` and released when it is dropped (or when the
+/// process dies — the OS drops the lock with the file descriptor, so a crash
+/// never leaves a stale lock behind; the empty lock file itself is harmless).
+#[derive(Debug)]
+pub(crate) struct StoreLock {
+    _file: File,
+}
+
+impl StoreLock {
+    pub(crate) fn lock_path(store: &Path) -> PathBuf {
+        store.with_extension("h5.lock")
+    }
+
+    pub(crate) fn acquire(store: &Path) -> Result<Self, MemoryError> {
+        let path = Self::lock_path(store);
+        let file = OpenOptions::new()
+            .create(true)
+            .truncate(false)
+            .write(true)
+            .open(&path)?;
+        // A previous owner may be mid-teardown (e.g. an `AsyncHDF5Memory`
+        // dropped without `shutdown()`: its background task releases the
+        // store a moment later), so give the lock a short, bounded grace
+        // period before reporting a genuine second writer.
+        let mut attempts_left = LOCK_RETRIES;
+        loop {
+            match file.try_lock() {
+                Ok(()) => return Ok(Self { _file: file }),
+                Err(TryLockError::WouldBlock) if attempts_left > 0 => {
+                    attempts_left -= 1;
+                    std::thread::sleep(LOCK_RETRY_DELAY);
+                }
+                Err(TryLockError::WouldBlock) => {
+                    return Err(MemoryError::Locked(format!(
+                        "{} is already open in this or another process (lock file {})",
+                        store.display(),
+                        path.display()
+                    )));
+                }
+                Err(TryLockError::Error(e)) => return Err(MemoryError::Io(e)),
+            }
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn second_acquire_fails_until_first_is_dropped() {
+        let dir = tempfile::TempDir::new().unwrap();
+        let store = dir.path().join("s.h5");
+        let first = StoreLock::acquire(&store).unwrap();
+        assert!(matches!(
+            StoreLock::acquire(&store),
+            Err(MemoryError::Locked(_))
+        ));
+        drop(first);
+        StoreLock::acquire(&store).unwrap();
+    }
+}
@@ -167,10 +167,17 @@ pub fn auto_select_strategy(num_vectors: usize, hw: &HardwareCapabilities) -> Se
 /// This dispatches to the appropriate search implementation based on the
 /// selected strategy. For IVF-PQ, an index must be provided externally
 /// (this function uses brute-force fallback if no IVF-PQ index is available).
+///
+/// `vectors_flat` is `vectors` flattened into one contiguous `[N × dim]`
+/// row-major buffer (e.g. `MemoryCache::embeddings_flat`, maintained
+/// incrementally alongside `vectors`). It's only consulted by the
+/// `Blas`/`Accelerate` strategies, which otherwise re-flatten the whole
+/// corpus on every call — passing the already-flat buffer skips that copy.
 #[allow(clippy::too_many_arguments)]
 pub fn search_with_metrics(
    query: &[f32],
    vectors: &[Vec<f32>],
+    vectors_flat: &[f32],
    norms: &[f32],
    tombstones: &[u8],
    k: usize,
@@ -178,6 +185,10 @@ pub fn search_with_metrics(
    #[cfg(feature = "gpu")] gpu_backend: Option<&crate::gpu_search::GpuSearchBackend>,
    #[cfg(not(feature = "gpu"))] _gpu_backend: Option<&()>,
 ) -> (Vec<(usize, f32)>, SearchMetrics) {
+    // Only read by the Blas/Accelerate arms below, which are themselves
+    // feature-gated — reference it unconditionally so a build with neither
+    // feature enabled doesn't warn about an unused parameter.
+    let _ = vectors_flat;
    let start = Instant::now();
    let active_count = tombstones.iter().filter(|&&t| t == 0).count();

@@ -197,7 +208,14 @@ pub fn search_with_metrics(
            gpu_active = false;
            #[cfg(feature = "fast-math")]
            {
-                crate::blas_search::blas_cosine_batch(query, vectors, norms, tombstones, k)
+                crate::blas_search::blas_cosine_batch_flat(
+                    query,
+                    vectors_flat,
+                    norms,
+                    tombstones,
+                    query.len(),
+                    k,
+                )
            }
            #[cfg(not(feature = "fast-math"))]
            {
@@ -211,8 +229,13 @@ pub fn search_with_metrics(
            gpu_active = false;
            #[cfg(any(feature = "accelerate", feature = "openblas"))]
            {
-                crate::accelerate_search::accelerate_cosine_batch_vecs(
-                    query, vectors, norms, tombstones, k,
+                crate::accelerate_search::accelerate_cosine_batch(
+                    query,
+                    vectors_flat,
+                    norms,
+                    tombstones,
+                    query.len(),
+                    k,
                )
            }
            #[cfg(not(any(feature = "accelerate", feature = "openblas")))]
@@ -325,6 +348,10 @@ mod tests {
        (0..n).map(|_| (0..dim).map(|_| next()).collect()).collect()
    }

+    fn flatten(vectors: &[Vec<f32>]) -> Vec<f32> {
+        vectors.iter().flatten().copied().collect()
+    }
+
    // --- auto_select_strategy tests ---

    #[test]
@@ -490,6 +517,7 @@ mod tests {
        let (results, metrics) = search_with_metrics(
            &query,
            &vectors,
+            &flatten(&vectors),
            &norms,
            &tombstones,
            5,
@@ -520,6 +548,7 @@ mod tests {
        let (results, metrics) = search_with_metrics(
            &query,
            &vectors,
+            &flatten(&vectors),
            &norms,
            &tombstones,
            10,
@@ -545,6 +574,7 @@ mod tests {
        let (_, metrics) = search_with_metrics(
            &query,
            &vectors,
+            &flatten(&vectors),
            &norms,
            &tombstones,
            10,
@@ -570,6 +600,7 @@ mod tests {
        let (results, _) = search_with_metrics(
            &query,
            &vectors,
+            &flatten(&vectors),
            &norms,
            &tombstones,
            10,
@@ -603,6 +634,7 @@ mod tests {
        let (results, metrics) = search_with_metrics(
            &query,
            &vectors,
+            &flatten(&vectors),
            &norms,
            &tombstones,
            100,
@@ -647,6 +679,7 @@ mod tests {
        let (_, metrics) = search_with_metrics(
            &query,
            &vectors,
+            &flatten(&vectors),
            &norms,
            &tombstones,
            5,
@@ -718,6 +751,7 @@ mod tests {
        let (results, metrics) = search_with_metrics(
            &query,
            &vectors,
+            &flatten(&vectors),
            &norms,
            &tombstones,
            10,
@@ -744,6 +778,7 @@ mod tests {
        let (results, metrics) = search_with_metrics(
            &query,
            &vectors,
+            &flatten(&vectors),
            &norms,
            &tombstones,
            10,
@@ -822,6 +857,7 @@ mod tests {
        let (results, metrics) = search_with_metrics(
            &query,
            &vectors,
+            &flatten(&vectors),
            &norms,
            &tombstones,
            10,
@@ -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));
@@ -13,16 +13,57 @@ use crate::MemoryError;

 const WAL_MAGIC: [u8; 4] = [0x45, 0x48, 0x57, 0x4C]; // "EHWL"

-/// Current WAL format version: every entry ends with a 4-byte CRC32 trailer
-/// (see [`TeeReader`]) so a bit-flip is detected and replay stops there
-/// instead of silently accepting corrupted data.
-const WAL_VERSION: u8 = 2;
+/// Bytes before the first entry: [`WAL_MAGIC`] (4) + version (1) + entry
+/// count (4). Named so the offset arithmetic in `open()` — which decides
+/// where an append lands, and therefore whether it is replayable — reads as
+/// a header length rather than a bare 9.
+const WAL_HEADER_LEN: u64 = WAL_MAGIC.len() as u64 + 1 + 4;

-/// The only other WAL version this crate still knows how to *read*: no
-/// per-entry CRC trailer. Written by versions of this crate before the CRC32
-/// hardening. `WalFile::open` migrates a legacy file to [`WAL_VERSION`] by
-/// recreating it fresh — safe because every real call site reads existing
-/// entries via [`WalFile::read_entries`] before calling `open` (see
+/// Current WAL format version: every entry's CRC32 trailer is computed over
+/// its own bytes *chained with the previous entry's stored CRC*
+/// (`crc32(entry_bytes ++ prev_crc.to_le_bytes())`, seeded with 0 for the
+/// first entry after a truncation). A per-entry CRC alone only detects a
+/// bit-flip within that entry; chaining additionally detects entries being
+/// reordered, duplicated, or spliced (e.g. a Tombstone moved before/after
+/// its target Save) — the moved/inserted entry's stored CRC was computed
+/// against a different predecessor than the one now in front of it on disk,
+/// so the chain breaks at that point and replay stops there.
+const WAL_VERSION: u8 = 4;
+
+/// The chained-CRC format before [`WalEntryType::Update`] records existed.
+/// Byte-for-byte the same framing as [`WAL_VERSION`], so it is read by the
+/// same code, and `WalFile::open` upgrades it in place by rewriting the
+/// header's version byte (the header is not covered by the CRC chain).
+///
+/// The bump exists for *older binaries*: they don't know record type 0x04,
+/// would treat it as a torn tail, and would truncate it — and everything
+/// after it — away. An unknown header version makes them refuse the file
+/// with a clear error instead.
+const WAL_VERSION_CHAINED_NO_UPDATE: u8 = 3;
+
+/// The previous WAL format version: still a CRC32 per entry (so a bit-flip
+/// within one entry is caught), but not chained to the previous entry's CRC
+/// (so reordering/splicing whole entries is not detected). Written by
+/// versions of this crate before the chaining hardening. Fully supported for
+/// reading via [`WalFile::read_entries`] — not restricted like
+/// [`WAL_VERSION_LEGACY_NO_CRC`], since it still verifies each entry
+/// individually. `WalFile::open` migrates it to [`WAL_VERSION`] by
+/// recreating the file fresh, the same as the legacy-no-CRC migration below.
+const WAL_VERSION_CRC_UNCHAINED: u8 = 2;
+
+/// The oldest WAL version this crate still knows how to *read*: no
+/// per-entry CRC trailer at all, so a bit-flip anywhere is silently
+/// accepted. Written by versions of this crate before the CRC32 hardening.
+/// Because of that — unlike [`WAL_VERSION_CRC_UNCHAINED`] — this version is
+/// deliberately *not* reachable through the public [`WalFile::read_entries`]
+/// API; only [`WalFile::read_entries_for_migration`] (used exclusively by
+/// `HDF5Memory::open`'s one-time migration path) will parse it. Flipping a
+/// version byte from 2/3 down to 1 no longer silently downgrades a file to
+/// the fully-unverified parser for an arbitrary caller.
+///
+/// `WalFile::open` migrates a legacy file to [`WAL_VERSION`] by recreating
+/// it fresh — safe because every real call site reads existing entries via
+/// [`WalFile::read_entries_for_migration`] before calling `open` (see
 /// `HDF5Memory::open`), so no data is lost.
 const WAL_VERSION_LEGACY_NO_CRC: u8 = 1;

@@ -37,6 +78,10 @@ pub enum WalEntryType {
    Save = 0x01,
    Tombstone = 0x02,
    ActivationUpdate = 0x03,
+    /// Replace the record at `update_index` in place (`save_or_update` hit).
+    /// Logged as a plain `Save` before this existed, so replay appended a
+    /// duplicate instead of updating.
+    Update = 0x04,
 }

 impl WalEntryType {
@@ -45,6 +90,7 @@ impl WalEntryType {
            0x01 => Some(Self::Save),
            0x02 => Some(Self::Tombstone),
            0x03 => Some(Self::ActivationUpdate),
+            0x04 => Some(Self::Update),
            _ => None,
        }
    }
@@ -61,6 +107,8 @@ pub struct WalEntry {
    pub tags: String,
    /// For tombstone entries: the index of the entry to delete.
    pub tombstone_index: Option<usize>,
+    /// For update entries: the index of the record to replace.
+    pub update_index: Option<usize>,
 }

 /// How many entries to accumulate before updating the header entry_count.
@@ -77,15 +125,77 @@ pub struct WalFile {
    entry_count: u32,
    /// Entries written since the last header count update.
    pending_header_sync: u32,
+    /// CRC32 chain state: the previous entry's stored CRC (0 if this file
+    /// has no entries yet), folded into the next entry's CRC computation.
+    /// Reset to 0 by `truncate()`/`create_fresh_wal_file`, and re-derived by
+    /// scanning existing entries when `open()` attaches to a non-empty file.
+    running_crc: u32,
+    /// Bytes of verified entries after the header (the length of the chain
+    /// `running_crc` covers). Together they form the [`WalMark`].
+    chain_len: u64,
+}
+
+/// What a WAL file's 9-byte header looks like, without reading any entries.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum WalHeaderStatus {
+    /// A version this build can read (current or legacy).
+    Readable,
+    /// Shorter than a header — e.g. a crash while the file was being created.
+    /// It cannot contain entries.
+    Torn,
+    /// Not a WAL file at all.
+    BadMagic,
+    /// Well-formed header from a version this build doesn't know — most
+    /// likely written by a *newer* build. Never discard this: the entries are
+    /// probably fine, this binary just can't read them.
+    UnknownVersion(u8),
+}
+
+/// Classify the header of the WAL at `path`.
+pub fn wal_header_status(path: &Path) -> std::io::Result<WalHeaderStatus> {
+    let mut header = [0u8; WAL_HEADER_LEN as usize];
+    let mut f = File::open(path)?;
+    let mut filled = 0;
+    while filled < header.len() {
+        match f.read(&mut header[filled..])? {
+            0 => return Ok(WalHeaderStatus::Torn),
+            n => filled += n,
+        }
+    }
+    if header[0..4] != WAL_MAGIC {
+        return Ok(WalHeaderStatus::BadMagic);
+    }
+    Ok(match header[4] {
+        WAL_VERSION
+        | WAL_VERSION_CHAINED_NO_UPDATE
+        | WAL_VERSION_CRC_UNCHAINED
+        | WAL_VERSION_LEGACY_NO_CRC => WalHeaderStatus::Readable,
+        v => WalHeaderStatus::UnknownVersion(v),
+    })
+}
+
+/// A position in a WAL's CRC chain: `len` bytes of entries after the header,
+/// whose chained CRC is `crc`.
+///
+/// A checkpoint stores the mark of the WAL prefix it folded into the `.h5`
+/// file. If the process dies after the new `.h5` is in place but before the
+/// WAL is truncated, the next `open()` finds that exact prefix still in the
+/// WAL and skips it instead of replaying it on top of data that already
+/// contains it (which used to duplicate every pending entry).
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub struct WalMark {
+    pub len: u64,
+    pub crc: u32,
 }

 impl WalFile {
    /// Open or create a WAL file. If it exists, read the header and entry count.
    ///
-    /// A legacy (pre-CRC) WAL file is migrated to the current format by
-    /// recreating it fresh — see [`WAL_VERSION_LEGACY_NO_CRC`]. Callers that
-    /// need the legacy file's entries must call [`WalFile::read_entries`]
-    /// first, before calling `open`.
+    /// A pre-chaining WAL file ([`WAL_VERSION_CRC_UNCHAINED`] or
+    /// [`WAL_VERSION_LEGACY_NO_CRC`]) is migrated to the current format by
+    /// recreating it fresh. Callers that need an existing file's entries must
+    /// call [`WalFile::read_entries`] (or, for a legacy-no-CRC file,
+    /// [`WalFile::read_entries_for_migration`]) first, before calling `open`.
    pub fn open(path: &Path) -> Result<Self, MemoryError> {
        if path.exists() {
            // Read existing header
@@ -102,20 +212,71 @@ impl WalFile {
            let mut ver = [0u8; 1];
            f.read_exact(&mut ver)?;
            match ver[0] {
-                WAL_VERSION => {
+                WAL_VERSION | WAL_VERSION_CHAINED_NO_UPDATE => {
+                    if ver[0] == WAL_VERSION_CHAINED_NO_UPDATE {
+                        // Same framing; stamp the current version so an older
+                        // binary refuses this file rather than truncating an
+                        // Update record it can't parse. See the constant.
+                        f.seek(SeekFrom::Start(4))?;
+                        f.write_all(&[WAL_VERSION])?;
+                        f.seek(SeekFrom::Start(5))?;
+                    }
                    let mut count_buf = [0u8; 4];
                    f.read_exact(&mut count_buf)?;
-                    let entry_count = u32::from_le_bytes(count_buf);
-                    // Seek to end for appending
-                    f.seek(SeekFrom::End(0))?;
+                    let header_count = u32::from_le_bytes(count_buf);
+                    // Scan any existing entries to resume the CRC chain
+                    // correctly for further appends (the header's count may
+                    // be stale from deferred group-commit sync, same
+                    // tolerance `read_entries` already has, so the scanned
+                    // count is also the more accurate of the two).
+                    let (entries, running_crc, verified_bytes) =
+                        read_chained_entries(&mut f, 0, None);
+                    let entry_count = if entries.is_empty() {
+                        header_count
+                    } else {
+                        entries.len() as u32
+                    };
+                    // Position the append at the end of the VERIFIED prefix,
+                    // and drop anything after it.
+                    //
+                    // This used to `seek(End(0))`, which appends PAST a torn
+                    // tail — the ordinary outcome of a crash mid-append. The
+                    // new entry is then chained to the last good entry, but
+                    // sits on disk behind the garbage:
+                    //
+                    //     [1..N verified][torn bytes][N+1 chained to N]
+                    //
+                    // Replay stops at the torn bytes, so N+1 is unreachable
+                    // FOREVER even though its `append` returned Ok and synced.
+                    // That is silent data loss in the one situation a WAL
+                    // exists for. Truncating to the verified end is the
+                    // standard recovery: the torn tail was never acknowledged
+                    // to any caller, so discarding it loses nothing, and the
+                    // chain then continues from a byte offset that matches
+                    // `running_crc`.
+                    let verified_end = WAL_HEADER_LEN + verified_bytes;
+                    let file_len = f.metadata()?.len();
+                    if file_len > verified_end {
+                        eprintln!(
+                            "clawhdf5-agent: WAL {} has {} unverifiable byte(s) after entry {}; \
+                             discarding them so appends stay replayable",
+                            path.display(),
+                            file_len - verified_end,
+                            entries.len()
+                        );
+                        f.set_len(verified_end)?;
+                    }
+                    f.seek(SeekFrom::Start(verified_end))?;
                    Ok(Self {
                        path: path.to_path_buf(),
                        file: Some(f),
                        entry_count,
                        pending_header_sync: 0,
+                        running_crc,
+                        chain_len: verified_bytes,
                    })
                }
-                WAL_VERSION_LEGACY_NO_CRC => {
+                WAL_VERSION_CRC_UNCHAINED | WAL_VERSION_LEGACY_NO_CRC => {
                    drop(f);
                    let f = create_fresh_wal_file(path)?;
                    Ok(Self {
@@ -123,6 +284,8 @@ impl WalFile {
                        file: Some(f),
                        entry_count: 0,
                        pending_header_sync: 0,
+                        running_crc: 0,
+                        chain_len: 0,
                    })
                }
                v => Err(MemoryError::Schema(format!("unsupported WAL version {v}"))),
@@ -134,6 +297,8 @@ impl WalFile {
                file: Some(f),
                entry_count: 0,
                pending_header_sync: 0,
+                running_crc: 0,
+                chain_len: 0,
            })
        }
    }
@@ -157,8 +322,20 @@ impl WalFile {
            4 + entry.session_id.len() +
            4 + entry.tags.len(),
        );
-        buf.push(WalEntryType::Save as u8);
-        buf.extend_from_slice(&entry.timestamp.to_le_bytes());
+        match entry.update_index {
+            Some(index) => {
+                let index = u32::try_from(index).map_err(|_| {
+                    MemoryError::Schema(format!("WAL update index {index} exceeds u32"))
+                })?;
+                buf.push(WalEntryType::Update as u8);
+                buf.extend_from_slice(&entry.timestamp.to_le_bytes());
+                buf.extend_from_slice(&index.to_le_bytes());
+            }
+            None => {
+                buf.push(WalEntryType::Save as u8);
+                buf.extend_from_slice(&entry.timestamp.to_le_bytes());
+            }
+        }
        serialize_str(&mut buf, &entry.chunk);
        buf.extend_from_slice(&(emb_len as u32).to_le_bytes());
        for &val in &entry.embedding {
@@ -168,7 +345,10 @@ impl WalFile {
        serialize_str(&mut buf, &entry.session_id);
        serialize_str(&mut buf, &entry.tags);

-        let crc = crc32(&buf);
+        // Chain this entry's CRC to the previous one's so reordering/
+        // splicing entries (not just flipping a bit within one) is detected
+        // on replay — see WAL_VERSION's doc comment.
+        let crc = chained_crc(&buf, self.running_crc);
        buf.extend_from_slice(&crc.to_le_bytes());

        let f = self
@@ -176,7 +356,9 @@ impl WalFile {
            .as_mut()
            .ok_or_else(|| MemoryError::Io(std::io::Error::other("WAL file not open")))?;
        f.write_all(&buf)?;
+        self.chain_len += buf.len() as u64;

+        self.running_crc = crc;
        self.entry_count += 1;
        self.pending_header_sync += 1;
        if self.pending_header_sync >= GROUP_COMMIT_SIZE {
@@ -191,7 +373,7 @@ impl WalFile {
        buf[0] = WalEntryType::Tombstone as u8;
        buf[1..9].copy_from_slice(&timestamp.to_le_bytes());
        buf[9..13].copy_from_slice(&(index as u32).to_le_bytes());
-        let crc = crc32(&buf[..13]);
+        let crc = chained_crc(&buf[..13], self.running_crc);
        buf[13..17].copy_from_slice(&crc.to_le_bytes());

        let f = self
@@ -199,7 +381,9 @@ impl WalFile {
            .as_mut()
            .ok_or_else(|| MemoryError::Io(std::io::Error::other("WAL file not open")))?;
        f.write_all(&buf)?;
+        self.chain_len += buf.len() as u64;

+        self.running_crc = crc;
        self.entry_count += 1;
        self.pending_header_sync += 1;
        if self.pending_header_sync >= GROUP_COMMIT_SIZE {
@@ -214,9 +398,46 @@ impl WalFile {
    /// (and may be stale if written with deferred group-commit updates). This
    /// tolerates both truncated files (crash mid-write) and stale header counts
    /// (crash before the next group-commit header sync). On a `WAL_VERSION`
-    /// file, a CRC32 mismatch on an entry is treated the same way — replay
-    /// stops there rather than accepting corrupted data.
+    /// file, a broken CRC chain (bit-flip, or an entry reordered/duplicated/
+    /// spliced in) is treated the same way — replay stops there rather than
+    /// accepting corrupted or tampered data. `WAL_VERSION_CRC_UNCHAINED`
+    /// files are read the same way minus the chain check (each entry's own
+    /// CRC is still verified).
+    ///
+    /// Does **not** read [`WAL_VERSION_LEGACY_NO_CRC`] files — that format has
+    /// no integrity verification at all, so it's only reachable through
+    /// [`WalFile::read_entries_for_migration`], used exclusively by
+    /// `HDF5Memory::open`'s one-time migration path. Calling this on a
+    /// legacy-no-CRC file returns a typed error instead of silently
+    /// downgrading to the unverified parser.
    pub fn read_entries(path: &Path) -> Result<Vec<WalEntry>, MemoryError> {
+        Self::read_entries_impl(path, false, None)
+    }
+
+    /// Like [`WalFile::read_entries`], but also accepts
+    /// [`WAL_VERSION_LEGACY_NO_CRC`] files (no per-entry integrity check at
+    /// all). Restricted to `pub(crate)` and named accordingly: the only
+    /// legitimate caller is `HDF5Memory::open`'s one-time migration of a
+    /// pre-CRC WAL file, which immediately recreates it in the current
+    /// format afterward. Do not use this for anything else.
+    ///
+    /// `applied` is the checkpoint mark read from the `.h5` file, if any: if
+    /// the WAL's chain passes through it (same byte length, same chained
+    /// CRC), everything up to that point is already in the `.h5` and is
+    /// dropped. If it never does — the normal case, because the WAL was
+    /// truncated after the checkpoint — every entry is returned.
+    pub(crate) fn read_entries_for_migration(
+        path: &Path,
+        applied: Option<WalMark>,
+    ) -> Result<Vec<WalEntry>, MemoryError> {
+        Self::read_entries_impl(path, true, applied)
+    }
+
+    fn read_entries_impl(
+        path: &Path,
+        allow_legacy_no_crc: bool,
+        applied: Option<WalMark>,
+    ) -> Result<Vec<WalEntry>, MemoryError> {
        if !path.exists() {
            return Ok(Vec::new());
        }
@@ -229,46 +450,62 @@ impl WalFile {
        }
        // entry_count is a pre-allocation hint only — we read until EOF.
        let entry_count_hint = u32::from_le_bytes([header[5], header[6], header[7], header[8]]);
-        let mut entries = Vec::with_capacity(entry_count_hint as usize);

        match header[4] {
-            WAL_VERSION => loop {
-                let raw_and_result = {
-                    let mut tee = TeeReader::new(&mut f);
-                    let result = read_one_entry(&mut tee);
-                    (tee.into_buf(), result)
-                };
-                let (raw, result) = raw_and_result;
-                let entry_opt = match result {
-                    Err(()) => break,
-                    Ok(v) => v,
-                };
-                let mut crc_buf = [0u8; 4];
-                if f.read_exact(&mut crc_buf).is_err() {
-                    break;
-                }
-                let stored_crc = u32::from_le_bytes(crc_buf);
-                if crc32(&raw) != stored_crc {
-                    // Corruption detected — stop replay here, same as a clean
-                    // truncation/EOF, rather than accepting the bad entry.
-                    break;
-                }
-                if let Some(entry) = entry_opt {
-                    entries.push(entry);
-                }
-            },
-            WAL_VERSION_LEGACY_NO_CRC => loop {
-                match read_one_entry(&mut f) {
-                    Err(()) => break,
-                    Ok(Some(entry)) => entries.push(entry),
-                    Ok(None) => {}
-                }
-            },
-            v => {
-                return Err(MemoryError::Schema(format!("unsupported WAL version {v}")));
+            WAL_VERSION | WAL_VERSION_CHAINED_NO_UPDATE => {
+                let (entries, _final_crc, _verified_bytes) =
+                    read_chained_entries(&mut f, 0, applied);
+                Ok(entries)
            }
+            WAL_VERSION_CRC_UNCHAINED => {
+                let mut entries = Vec::with_capacity(entry_count_hint as usize);
+                loop {
+                    let raw_and_result = {
+                        let mut tee = TeeReader::new(&mut f);
+                        let result = read_one_entry(&mut tee);
+                        (tee.into_buf(), result)
+                    };
+                    let (raw, result) = raw_and_result;
+                    let entry_opt = match result {
+                        Err(()) => break,
+                        Ok(v) => v,
+                    };
+                    let mut crc_buf = [0u8; 4];
+                    if f.read_exact(&mut crc_buf).is_err() {
+                        break;
+                    }
+                    let stored_crc = u32::from_le_bytes(crc_buf);
+                    if crc32(&raw) != stored_crc {
+                        // Corruption detected — stop replay here, same as a
+                        // clean truncation/EOF, rather than accepting the bad
+                        // entry.
+                        break;
+                    }
+                    if let Some(entry) = entry_opt {
+                        entries.push(entry);
+                    }
+                }
+                Ok(entries)
+            }
+            WAL_VERSION_LEGACY_NO_CRC if allow_legacy_no_crc => {
+                let mut entries = Vec::with_capacity(entry_count_hint as usize);
+                loop {
+                    match read_one_entry(&mut f) {
+                        Err(()) => break,
+                        Ok(Some(entry)) => entries.push(entry),
+                        Ok(None) => {}
+                    }
+                }
+                Ok(entries)
+            }
+            WAL_VERSION_LEGACY_NO_CRC => Err(MemoryError::Schema(
+                "WAL file is in the legacy no-CRC format (version 1), which read_entries() no \
+                 longer accepts — it has no per-entry integrity verification. Only the one-time \
+                 migration path (WalFile::open) can read and upgrade it."
+                    .into(),
+            )),
+            v => Err(MemoryError::Schema(format!("unsupported WAL version {v}"))),
        }
-        Ok(entries)
    }

    /// Truncate the WAL (after merge into .h5).
@@ -279,9 +516,20 @@ impl WalFile {
        self.file = Some(f);
        self.entry_count = 0;
        self.pending_header_sync = 0;
+        self.running_crc = 0;
+        self.chain_len = 0;
        Ok(())
    }

+    /// The mark covering every entry currently in this WAL. Store it with a
+    /// checkpoint taken from the state those entries produced.
+    pub fn mark(&self) -> WalMark {
+        WalMark {
+            len: self.chain_len,
+            crc: self.running_crc,
+        }
+    }
+
    /// Number of pending entries.
    pub fn pending_count(&self) -> u32 {
        self.entry_count
@@ -321,6 +569,28 @@ pub fn replay_into_cache(entries: &[WalEntry], cache: &mut crate::cache::MemoryC
                    entry.tags.clone(),
                );
            }
+            WalEntryType::Update => match entry.update_index {
+                // The index was valid when the record was written; if the
+                // store no longer has it, keep the data rather than drop it.
+                Some(idx) if idx < cache.len() => cache.update(
+                    idx,
+                    entry.chunk.clone(),
+                    entry.embedding.clone(),
+                    entry.source_channel.clone(),
+                    entry.timestamp,
+                    entry.session_id.clone(),
+                ),
+                _ => {
+                    cache.push(
+                        entry.chunk.clone(),
+                        entry.embedding.clone(),
+                        entry.source_channel.clone(),
+                        entry.timestamp,
+                        entry.session_id.clone(),
+                        entry.tags.clone(),
+                    );
+                }
+            },
            WalEntryType::Tombstone => {
                if let Some(idx) = entry.tombstone_index {
                    cache.mark_deleted(idx);
@@ -373,6 +643,81 @@ fn read_embedding<R: Read>(f: &mut R) -> Result<Vec<f32>, MemoryError> {
    Ok(vals)
 }

+/// Compute the CRC32 trailer for a `WAL_VERSION` entry, chaining in the
+/// previous entry's stored CRC (0 for the first entry after a truncation).
+fn chained_crc(entry_bytes: &[u8], prev_crc: u32) -> u32 {
+    let mut chained = Vec::with_capacity(entry_bytes.len() + 4);
+    chained.extend_from_slice(entry_bytes);
+    chained.extend_from_slice(&prev_crc.to_le_bytes());
+    crc32(&chained)
+}
+
+/// Read and verify all entries from a `WAL_VERSION` (chained-CRC) stream
+/// starting at the reader's current position, given the chain state to
+/// resume from (0 for a stream starting at the beginning of a fresh WAL).
+///
+/// Returns the parsed entries, the final running CRC — the chain state to
+/// continue from for further appends — and the number of BYTES consumed by
+/// those verified entries. Stops (without erroring) at the first entry that
+/// fails to parse or whose stored CRC doesn't match the expected chain value
+/// — a bit-flip, truncation/EOF, or an entry having been
+/// reordered/duplicated/spliced all produce a chain mismatch at that point,
+/// and are all handled the same way: replay stops there.
+///
+/// The byte count is what lets `open()` position an append at the end of the
+/// VERIFIED prefix rather than at end-of-file. Appending past a torn tail
+/// writes entries that replay can never reach — see `open`.
+///
+/// `applied`, when given, is a checkpoint mark: once the chain reaches exactly
+/// that position, the entries collected so far are discarded (they are
+/// already in the `.h5` file). A zero-length mark matches nothing.
+fn read_chained_entries<R: Read>(
+    f: &mut R,
+    start_crc: u32,
+    applied: Option<WalMark>,
+) -> (Vec<WalEntry>, u32, u64) {
+    let applied = applied.filter(|m| m.len > 0);
+    let mut entries = Vec::new();
+    let mut running_crc = start_crc;
+    let mut verified_bytes: u64 = 0;
+    loop {
+        let raw_and_result = {
+            let mut tee = TeeReader::new(f);
+            let result = read_one_entry(&mut tee);
+            (tee.into_buf(), result)
+        };
+        let (raw, result) = raw_and_result;
+        let entry_opt = match result {
+            Err(()) => break,
+            Ok(v) => v,
+        };
+        let mut crc_buf = [0u8; 4];
+        if f.read_exact(&mut crc_buf).is_err() {
+            break;
+        }
+        let stored_crc = u32::from_le_bytes(crc_buf);
+        if chained_crc(&raw, running_crc) != stored_crc {
+            break;
+        }
+        running_crc = stored_crc;
+        // Only counted once the entry AND its CRC trailer verified, so the
+        // offset always points just past a complete, checked entry.
+        verified_bytes += raw.len() as u64 + crc_buf.len() as u64;
+        if let Some(entry) = entry_opt {
+            entries.push(entry);
+        }
+        if applied
+            == Some(WalMark {
+                len: verified_bytes,
+                crc: running_crc,
+            })
+        {
+            entries.clear();
+        }
+    }
+    (entries, running_crc, verified_bytes)
+}
+
 /// Create a fresh WAL file at `path` with the current-version header,
 /// truncating/overwriting anything already there.
 fn create_fresh_wal_file(path: &Path) -> Result<File, MemoryError> {
@@ -430,7 +775,14 @@ fn read_one_entry<R: Read>(r: &mut R) -> Result<Option<WalEntry>, ()> {
    let timestamp = f64::from_le_bytes(ts_buf);

    match entry_type {
-        WalEntryType::Save => {
+        WalEntryType::Save | WalEntryType::Update => {
+            let update_index = if entry_type == WalEntryType::Update {
+                let mut idx_buf = [0u8; 4];
+                r.read_exact(&mut idx_buf).map_err(|_| ())?;
+                Some(u32::from_le_bytes(idx_buf) as usize)
+            } else {
+                None
+            };
            let chunk = read_len_prefixed_str(r).map_err(|_| ())?;
            let embedding = read_embedding(r).map_err(|_| ())?;
            let source_channel = read_len_prefixed_str(r).map_err(|_| ())?;
@@ -445,6 +797,7 @@ fn read_one_entry<R: Read>(r: &mut R) -> Result<Option<WalEntry>, ()> {
                session_id,
                tags,
                tombstone_index: None,
+                update_index,
            }))
        }
        WalEntryType::Tombstone => {
@@ -460,6 +813,7 @@ fn read_one_entry<R: Read>(r: &mut R) -> Result<Option<WalEntry>, ()> {
                session_id: String::new(),
                tags: String::new(),
                tombstone_index: Some(idx),
+                update_index: None,
            }))
        }
        WalEntryType::ActivationUpdate => Ok(None),
@@ -483,6 +837,7 @@ mod tests {
            session_id: "sess-001".to_string(),
            tags: "tag1,tag2".to_string(),
            tombstone_index: None,
+            update_index: None,
        }
    }

@@ -601,7 +956,7 @@ mod tests {
        let dir = TempDir::new().unwrap();
        let wal_path = dir.path().join("test.h5.wal");
        let unicode_chunk = "Hello 世界! 🌍 émojis & ünïcödé";
-        let embedding = vec![0.1, -0.2, 3.14159, f32::MAX, f32::MIN_POSITIVE];
+        let embedding = vec![0.1, -0.2, 3.4567, f32::MAX, f32::MIN_POSITIVE];
        {
            let mut wal = WalFile::open(&wal_path).unwrap();
            let entry = WalEntry {
@@ -613,6 +968,7 @@ mod tests {
                session_id: "sess-öö-123".to_string(),
                tags: "α,β,γ".to_string(),
                tombstone_index: None,
+                update_index: None,
            };
            wal.append_save(&entry).unwrap();
        }
@@ -747,6 +1103,148 @@ mod tests {
        assert!(entries.is_empty());
    }

+    /// Reopen `path` and return the stored chunks in order.
+    fn reopen_chunks(path: &std::path::Path) -> Vec<String> {
+        let mem = HDF5Memory::open(path).unwrap();
+        mem.cache.chunks.clone()
+    }
+
+    #[test]
+    fn crash_between_checkpoint_and_wal_truncate_does_not_duplicate() {
+        // flush() writes the new .h5 and only then truncates the WAL. Dying in
+        // between leaves BOTH a .h5 that contains the pending entries and a
+        // WAL that still lists them; replaying blindly used to double them.
+        let dir = TempDir::new().unwrap();
+        let config = make_config(&dir);
+        let h5_path = config.path.clone();
+        let wal_path = h5_path.with_extension("h5.wal");
+        let stale_wal = dir.path().join("stale.wal");
+
+        {
+            let mut mem = HDF5Memory::create(config).unwrap();
+            for name in ["a", "b", "c"] {
+                mem.save(make_entry(name, &[1.0, 0.0, 0.0, 0.0])).unwrap();
+            }
+            assert_eq!(mem.wal_pending_count(), 3);
+            std::fs::copy(&wal_path, &stale_wal).unwrap();
+            mem.flush_wal().unwrap();
+        }
+        // Undo the truncate: this is the on-disk state right after the crash.
+        std::fs::copy(&stale_wal, &wal_path).unwrap();
+        assert_eq!(WalFile::read_entries(&wal_path).unwrap().len(), 3);
+
+        assert_eq!(reopen_chunks(&h5_path), ["a", "b", "c"]);
+
+        // Entries appended to that same WAL after recovery are still replayed.
+        {
+            let mut mem = HDF5Memory::open(&h5_path).unwrap();
+            mem.save(make_entry("d", &[0.0, 1.0, 0.0, 0.0])).unwrap();
+        }
+        assert_eq!(reopen_chunks(&h5_path), ["a", "b", "c", "d"]);
+    }
+
+    #[test]
+    fn entries_written_after_a_completed_checkpoint_are_all_replayed() {
+        // Normal case: the checkpoint's mark refers to a WAL that has since
+        // been truncated, so it must not suppress anything in the new one —
+        // including when the new WAL grows past the old mark's length.
+        let dir = TempDir::new().unwrap();
+        let config = make_config(&dir);
+        let h5_path = config.path.clone();
+        {
+            let mut mem = HDF5Memory::create(config).unwrap();
+            mem.save(make_entry("a", &[1.0, 0.0, 0.0, 0.0])).unwrap();
+            mem.flush_wal().unwrap();
+            for name in ["b", "c", "d"] {
+                mem.save(make_entry(name, &[1.0, 0.0, 0.0, 0.0])).unwrap();
+            }
+        }
+        assert_eq!(reopen_chunks(&h5_path), ["a", "b", "c", "d"]);
+    }
+
+    #[test]
+    fn save_or_update_replays_as_update_not_duplicate() {
+        let dir = TempDir::new().unwrap();
+        let config = make_config(&dir);
+        let h5_path = config.path.clone();
+        {
+            let mut mem = HDF5Memory::create(config).unwrap();
+            let mut first = make_entry("v1", &[1.0, 0.0, 0.0, 0.0]);
+            first.tags = "key".into();
+            let mut second = make_entry("v2", &[0.0, 1.0, 0.0, 0.0]);
+            second.tags = "key".into();
+            let a = mem.save_or_update(first).unwrap();
+            mem.save(make_entry("other", &[0.0, 0.0, 1.0, 0.0]))
+                .unwrap();
+            let b = mem.save_or_update(second).unwrap();
+            assert_eq!(a, b);
+            assert_eq!(mem.cache.chunks, ["v2", "other"]);
+            // Dropped without a checkpoint: all three records live in the WAL.
+        }
+        let mem = HDF5Memory::open(&h5_path).unwrap();
+        assert_eq!(mem.cache.chunks, ["v2", "other"]);
+        assert_eq!(mem.cache.embeddings[0], [0.0, 1.0, 0.0, 0.0]);
+    }
+
+    #[test]
+    fn v3_wal_is_read_and_upgraded_in_place() {
+        let dir = TempDir::new().unwrap();
+        let wal_path = dir.path().join("old.wal");
+        {
+            let mut wal = WalFile::open(&wal_path).unwrap();
+            wal.append_save(&make_wal_entry("kept", &[1.0])).unwrap();
+        }
+        // Rewrite the header as the pre-Update chained format.
+        let mut bytes = std::fs::read(&wal_path).unwrap();
+        bytes[4] = WAL_VERSION_CHAINED_NO_UPDATE;
+        std::fs::write(&wal_path, &bytes).unwrap();
+
+        assert_eq!(WalFile::read_entries(&wal_path).unwrap().len(), 1);
+        {
+            let mut wal = WalFile::open(&wal_path).unwrap();
+            assert_eq!(wal.pending_count(), 1);
+            wal.append_save(&make_wal_entry("new", &[2.0])).unwrap();
+        }
+        assert_eq!(std::fs::read(&wal_path).unwrap()[4], WAL_VERSION);
+        let chunks: Vec<_> = WalFile::read_entries(&wal_path)
+            .unwrap()
+            .into_iter()
+            .map(|e| e.chunk)
+            .collect();
+        assert_eq!(chunks, ["kept", "new"]);
+    }
+
+    #[test]
+    fn mark_matching_is_exact() {
+        let dir = TempDir::new().unwrap();
+        let wal_path = dir.path().join("m.wal");
+        let mut wal = WalFile::open(&wal_path).unwrap();
+        wal.append_save(&make_wal_entry("one", &[1.0])).unwrap();
+        let after_one = wal.mark();
+        wal.append_save(&make_wal_entry("two", &[2.0])).unwrap();
+        let after_two = wal.mark();
+        drop(wal);
+
+        let read = |m| {
+            WalFile::read_entries_for_migration(&wal_path, m)
+                .unwrap()
+                .into_iter()
+                .map(|e| e.chunk)
+                .collect::<Vec<_>>()
+        };
+        assert_eq!(read(None), ["one", "two"]);
+        assert_eq!(read(Some(after_one)), ["two"]);
+        assert!(read(Some(after_two)).is_empty());
+        // Right length, wrong CRC (a different WAL generation): skip nothing.
+        let foreign = WalMark {
+            crc: after_one.crc ^ 1,
+            ..after_one
+        };
+        assert_eq!(read(Some(foreign)), ["one", "two"]);
+        // Reopening resumes the same mark.
+        assert_eq!(WalFile::open(&wal_path).unwrap().mark(), after_two);
+    }
+
    #[test]
    fn test_wal_replay_on_open() {
        // Test WAL replay using read_entries + replay_into_cache directly,
@@ -912,16 +1410,157 @@ mod tests {
        assert_eq!(entries[0].chunk, "first");
    }

+    /// A crash mid-append leaves a torn final entry. Reopening the WAL must
+    /// place the next append at the end of the VERIFIED prefix, not at
+    /// end-of-file, or that append is written behind garbage the replay
+    /// scanner stops at — unreachable forever despite having returned Ok.
+    ///
+    /// This is the ordinary crash case, so getting it wrong loses
+    /// acknowledged writes in exactly the situation a WAL exists for.
    #[test]
-    fn test_wal_reads_legacy_v1_format_without_crc() {
+    fn test_wal_append_after_torn_tail_stays_replayable() {
        let dir = TempDir::new().unwrap();
-        let wal_path = dir.path().join("legacy.h5.wal");
+        let wal_path = dir.path().join("test.h5.wal");
+
+        let mut wal = WalFile::open(&wal_path).unwrap();
+        wal.append_save(&make_wal_entry("first", &[1.0, 2.0]))
+            .unwrap();
+        drop(wal);
+
+        // Simulate the crash: a partial entry appended after the good one.
+        {
+            use std::io::Write;
+            let mut f = std::fs::OpenOptions::new()
+                .append(true)
+                .open(&wal_path)
+                .unwrap();
+            f.write_all(&[0xAB, 0xCD, 0xEF, 0x01, 0x02]).unwrap();
+            f.flush().unwrap();
+        }
+
+        // Reopen and append. The torn bytes must not survive between the
+        // verified prefix and the new entry.
+        let mut wal = WalFile::open(&wal_path).unwrap();
+        wal.append_save(&make_wal_entry("second", &[3.0, 4.0]))
+            .unwrap();
+        drop(wal);
+
+        let entries = WalFile::read_entries(&wal_path).unwrap();
+        assert_eq!(
+            entries.len(),
+            2,
+            "the append after a torn tail must be replayable; got {} entr(y/ies) — \
+             the post-crash write was silently lost",
+            entries.len()
+        );
+    }
+
+    /// Reordering two entries on disk must break the CRC chain — the
+    /// second entry's stored CRC was computed against the first entry's
+    /// real CRC, not against the chain state a reader sees after swapping
+    /// them, so replay stops immediately instead of accepting the tampered
+    /// order (INT-09).
+    #[test]
+    fn test_wal_detects_reordered_entries() {
+        let dir = TempDir::new().unwrap();
+        let wal_path = dir.path().join("test.h5.wal");
+        let mut wal = WalFile::open(&wal_path).unwrap();
+        wal.append_save(&make_wal_entry("first", &[1.0, 2.0]))
+            .unwrap();
+        let len_after_first = std::fs::metadata(&wal_path).unwrap().len() as usize;
+        wal.append_save(&make_wal_entry("second", &[3.0, 4.0]))
+            .unwrap();
+        let len_after_second = std::fs::metadata(&wal_path).unwrap().len() as usize;
+        drop(wal);
+
+        let bytes = std::fs::read(&wal_path).unwrap();
+        let header_len = 9usize;
+        let entry1_bytes = bytes[header_len..len_after_first].to_vec();
+        let entry2_bytes = bytes[len_after_first..len_after_second].to_vec();
+
+        let mut spliced = bytes[..header_len].to_vec();
+        spliced.extend_from_slice(&entry2_bytes);
+        spliced.extend_from_slice(&entry1_bytes);
+        std::fs::write(&wal_path, &spliced).unwrap();
+
+        let entries = WalFile::read_entries(&wal_path).unwrap();
+        assert!(
+            entries.is_empty(),
+            "reordered entries must break the CRC chain and stop replay, got {} entries",
+            entries.len()
+        );
+    }
+
+    /// Splicing a third-party entry in between two legitimate entries (e.g.
+    /// moving a Tombstone in front of the Save it's meant to follow) must
+    /// also break the chain for everything after the splice point.
+    #[test]
+    fn test_wal_detects_spliced_entry() {
+        let dir = TempDir::new().unwrap();
+        let wal_path = dir.path().join("test.h5.wal");
+        let mut wal = WalFile::open(&wal_path).unwrap();
+        wal.append_save(&make_wal_entry("first", &[1.0])).unwrap();
+        let len_after_first = std::fs::metadata(&wal_path).unwrap().len() as usize;
+        wal.append_save(&make_wal_entry("second", &[2.0])).unwrap();
+        let len_after_second = std::fs::metadata(&wal_path).unwrap().len() as usize;
+        wal.append_save(&make_wal_entry("third", &[3.0])).unwrap();
+        drop(wal);
+
+        let bytes = std::fs::read(&wal_path).unwrap();
+        let entry2_bytes = bytes[len_after_first..len_after_second].to_vec();
+
+        // Duplicate "second" right after itself: [first][second][second][third]
+        let mut spliced = bytes[..len_after_second].to_vec();
+        spliced.extend_from_slice(&entry2_bytes);
+        spliced.extend_from_slice(&bytes[len_after_second..]);
+        std::fs::write(&wal_path, &spliced).unwrap();
+
+        let entries = WalFile::read_entries(&wal_path).unwrap();
+        assert_eq!(
+            entries.len(),
+            2,
+            "replay must stop at the spliced duplicate, keeping only the entries before it"
+        );
+        assert_eq!(entries[0].chunk, "first");
+        assert_eq!(entries[1].chunk, "second");
+    }
+
+    /// A WAL closed (without truncating) and reopened must continue the CRC
+    /// chain correctly for newly appended entries — this is the normal
+    /// crash-restart-without-flush scenario (`HDF5Memory::open` replays
+    /// existing entries, then reopens the same file for further appends
+    /// without clearing it), and must not produce a false "reordering"
+    /// detection for its own legitimately-appended entries.
+    #[test]
+    fn test_wal_chain_continues_across_reopen() {
+        let dir = TempDir::new().unwrap();
+        let wal_path = dir.path().join("test.h5.wal");
+
+        let mut wal = WalFile::open(&wal_path).unwrap();
+        wal.append_save(&make_wal_entry("first", &[1.0])).unwrap();
+        drop(wal); // simulate a restart without ever truncating the WAL
+
+        let mut wal2 = WalFile::open(&wal_path).unwrap();
+        wal2.append_save(&make_wal_entry("second", &[2.0])).unwrap();
+        drop(wal2);
+
+        let entries = WalFile::read_entries(&wal_path).unwrap();
+        assert_eq!(
+            entries.len(),
+            2,
+            "both pre- and post-reopen entries must replay cleanly"
+        );
+        assert_eq!(entries[0].chunk, "first");
+        assert_eq!(entries[1].chunk, "second");
+    }
+
+    /// Build a legacy (WAL_VERSION_LEGACY_NO_CRC) WAL file containing one
+    /// Save entry, with no trailing CRC32.
+    fn build_legacy_v1_wal_bytes() -> Vec<u8> {
        let mut buf = Vec::new();
        buf.extend_from_slice(&WAL_MAGIC);
        buf.push(WAL_VERSION_LEGACY_NO_CRC);
        buf.extend_from_slice(&1u32.to_le_bytes());
-        // One Save entry in the old format: type + timestamp + fields, with
-        // no trailing CRC32.
        buf.push(WalEntryType::Save as u8);
        buf.extend_from_slice(&42.0f64.to_le_bytes());
        serialize_str(&mut buf, "legacy-chunk");
@@ -933,14 +1572,39 @@ mod tests {
        serialize_str(&mut buf, "chan");
        serialize_str(&mut buf, "sess");
        serialize_str(&mut buf, "tags");
-        std::fs::write(&wal_path, &buf).unwrap();
+        buf
+    }

-        let entries = WalFile::read_entries(&wal_path).unwrap();
+    #[test]
+    fn test_wal_reads_legacy_v1_format_without_crc() {
+        let dir = TempDir::new().unwrap();
+        let wal_path = dir.path().join("legacy.h5.wal");
+        std::fs::write(&wal_path, build_legacy_v1_wal_bytes()).unwrap();
+
+        // Only the migration-only reader may read a legacy no-CRC file.
+        let entries = WalFile::read_entries_for_migration(&wal_path, None).unwrap();
        assert_eq!(entries.len(), 1);
        assert_eq!(entries[0].chunk, "legacy-chunk");
        assert_eq!(entries[0].embedding, vec![1.0, 2.0]);
    }

+    /// The public `read_entries` must reject a legacy no-CRC file instead of
+    /// silently downgrading to the fully-unverified parser (INT-09) — flipping
+    /// a version byte from 2/3 down to 1 must not be a way to bypass every
+    /// integrity check for an arbitrary caller of the public API.
+    #[test]
+    fn test_wal_read_entries_rejects_legacy_v1_format() {
+        let dir = TempDir::new().unwrap();
+        let wal_path = dir.path().join("legacy.h5.wal");
+        std::fs::write(&wal_path, build_legacy_v1_wal_bytes()).unwrap();
+
+        let result = WalFile::read_entries(&wal_path);
+        assert!(
+            result.is_err(),
+            "read_entries() must reject a legacy no-CRC WAL file, not silently parse it"
+        );
+    }
+
    #[test]
    fn test_wal_open_migrates_legacy_v1_to_current_version() {
        let dir = TempDir::new().unwrap();
@@ -0,0 +1,187 @@
+//! Crash-recovery matrix for `HDF5Memory`.
+//!
+//! A process crash leaves whatever reached the OS on disk. These tests build
+//! the on-disk images such a crash can leave behind — after every operation,
+//! inside the checkpoint window (new `.h5` in place, WAL not yet truncated),
+//! and with the WAL torn at every possible length — then reopen each image
+//! and check the recovered store against a model of what was acknowledged.
+//!
+//! Invariants:
+//! * never a duplicated or invented record;
+//! * an image taken between operations recovers *exactly* the acknowledged
+//!   state;
+//! * a torn WAL recovers the last checkpoint plus a prefix of the operations
+//!   logged since.
+
+use std::path::{Path, PathBuf};
+
+use clawhdf5_agent::{AgentMemory, HDF5Memory, MemoryConfig, MemoryEntry};
+use tempfile::TempDir;
+
+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 below(&mut self, n: usize) -> usize {
+        (self.next() % n.max(1) as u64) as usize
+    }
+}
+
+fn entry(chunk: &str, tags: &str) -> MemoryEntry {
+    MemoryEntry {
+        chunk: chunk.to_string(),
+        embedding: vec![1.0, 0.0, 0.0, 0.0],
+        source_channel: "test".into(),
+        timestamp: 1.0,
+        session_id: "s".into(),
+        tags: tags.to_string(),
+    }
+}
+
+fn wal_path(h5: &Path) -> PathBuf {
+    h5.with_extension("h5.wal")
+}
+
+/// Copy the store (`.h5` + WAL) into a fresh directory, as a crash image.
+fn image(h5: &Path, into: &TempDir, name: &str) -> PathBuf {
+    let dest = into.path().join(format!("{name}.h5"));
+    std::fs::copy(h5, &dest).unwrap();
+    if wal_path(h5).exists() {
+        std::fs::copy(wal_path(h5), wal_path(&dest)).unwrap();
+    }
+    dest
+}
+
+fn recovered(h5: &Path) -> Vec<String> {
+    // Read-only: the image must not be modified, and no lock is needed.
+    HDF5Memory::open_read_only(h5).unwrap().cache.chunks.clone()
+}
+
+/// Apply one random operation to the store and to the model.
+fn step(mem: &mut HDF5Memory, model: &mut Vec<String>, rng: &mut Rng, n: usize) {
+    match rng.below(6) {
+        0 => mem.flush_wal().unwrap(),
+        1 if !model.is_empty() => {
+            // Update an existing record in place, addressed by its tag.
+            let idx = rng.below(model.len());
+            let chunk = format!("u{n}");
+            assert_eq!(
+                mem.save_or_update(entry(&chunk, &format!("tag{idx}")))
+                    .unwrap(),
+                idx
+            );
+            model[idx] = chunk;
+        }
+        _ => {
+            let chunk = format!("c{n}");
+            mem.save(entry(&chunk, &format!("tag{}", model.len())))
+                .unwrap();
+            model.push(chunk);
+        }
+    }
+}
+
+#[test]
+fn image_after_every_operation_recovers_the_acknowledged_state() {
+    for seed in 0..40u64 {
+        let mut rng = Rng(seed);
+        let dir = TempDir::new().unwrap();
+        let images = TempDir::new().unwrap();
+        let mut config = MemoryConfig::new(dir.path().join("store.h5"), "agent", 4);
+        config.wal_enabled = true;
+        config.wal_max_entries = 1 + rng.below(6); // force frequent checkpoints
+        let h5 = config.path.clone();
+        let mut mem = HDF5Memory::create(config).unwrap();
+        let mut model = Vec::new();
+
+        for n in 0..30 {
+            step(&mut mem, &mut model, &mut rng, n);
+            let img = image(&h5, &images, &format!("s{seed}-{n}"));
+            assert_eq!(recovered(&img), model, "seed {seed}, after op {n}");
+        }
+    }
+}
+
+#[test]
+fn crash_inside_the_checkpoint_window_never_duplicates() {
+    for seed in 0..40u64 {
+        let mut rng = Rng(seed ^ 0xABCD);
+        let dir = TempDir::new().unwrap();
+        let images = TempDir::new().unwrap();
+        let mut config = MemoryConfig::new(dir.path().join("store.h5"), "agent", 4);
+        config.wal_enabled = true;
+        config.wal_max_entries = 1000; // checkpoints only when we ask
+        let h5 = config.path.clone();
+        let mut mem = HDF5Memory::create(config).unwrap();
+        let mut model = Vec::new();
+
+        for round in 0..4 {
+            for n in 0..(1 + rng.below(6)) {
+                step(&mut mem, &mut model, &mut rng, round * 100 + n);
+            }
+            // The WAL as it is just before the checkpoint...
+            let stale_wal = images.path().join(format!("stale-{seed}-{round}.wal"));
+            if wal_path(&h5).exists() {
+                std::fs::copy(wal_path(&h5), &stale_wal).unwrap();
+            }
+            mem.flush_wal().unwrap();
+            // ...put back next to the NEW .h5: the crash-in-the-window image.
+            let img = image(&h5, &images, &format!("w{seed}-{round}"));
+            if stale_wal.exists() {
+                std::fs::copy(&stale_wal, wal_path(&img)).unwrap();
+            }
+            assert_eq!(recovered(&img), model, "seed {seed}, round {round}");
+        }
+    }
+}
+
+#[test]
+fn torn_wal_recovers_checkpoint_plus_a_prefix() {
+    let dir = TempDir::new().unwrap();
+    let images = TempDir::new().unwrap();
+    let mut config = MemoryConfig::new(dir.path().join("store.h5"), "agent", 4);
+    config.wal_enabled = true;
+    config.wal_max_entries = 1000;
+    let h5 = config.path.clone();
+    let mut mem = HDF5Memory::create(config).unwrap();
+
+    for name in ["a", "b"] {
+        mem.save(entry(name, name)).unwrap();
+    }
+    mem.flush_wal().unwrap();
+    let checkpointed = vec!["a".to_string(), "b".to_string()];
+
+    // States the store passes through as each later op is logged.
+    let mut states = vec![checkpointed.clone()];
+    let mut model = checkpointed.clone();
+    mem.save(entry("c", "c")).unwrap();
+    model.push("c".into());
+    states.push(model.clone());
+    mem.save_or_update(entry("a2", "a")).unwrap();
+    model[0] = "a2".into();
+    states.push(model.clone());
+    mem.save(entry("d", "d")).unwrap();
+    model.push("d".into());
+    states.push(model.clone());
+
+    let full_wal = std::fs::read(wal_path(&h5)).unwrap();
+    let mut seen = std::collections::BTreeSet::new();
+    for len in 0..=full_wal.len() {
+        let img = image(&h5, &images, &format!("t{len}"));
+        std::fs::write(wal_path(&img), &full_wal[..len]).unwrap();
+        let got = recovered(&img);
+        let which = states
+            .iter()
+            .position(|s| *s == got)
+            .unwrap_or_else(|| panic!("WAL torn at {len} bytes recovered {got:?}"));
+        seen.insert(which);
+    }
+    // Every intermediate state is reachable, and the full WAL gives the last.
+    assert_eq!(seen.into_iter().collect::<Vec<_>>(), [0, 1, 2, 3]);
+}
@@ -196,7 +196,7 @@ fn test_migration_round_trip() {
    mem.add_relation(e1, e2, "discusses", 0.8).unwrap();

    // Verify all data transferred by reopening
-    let reopened = HDF5Memory::open(&path).unwrap();
+    let reopened = HDF5Memory::open_read_only(&path).unwrap();
    assert_eq!(reopened.count(), 500);

    // Verify sessions
@@ -266,7 +266,7 @@ fn test_knowledge_graph_workflow() {
    assert_eq!(entity.entity_type, "library");

    // Persistence
-    let reopened = HDF5Memory::open(&path).unwrap();
+    let reopened = HDF5Memory::open_read_only(&path).unwrap();
    assert_eq!(reopened.knowledge().entities.len(), 4);
    assert_eq!(reopened.knowledge().relations.len(), 4);

@@ -316,7 +316,7 @@ fn test_multi_session_workflow() {
    assert_eq!(mem.count(), 100); // 5 sessions * 20 entries

    // Reopen and verify sessions
-    let reopened = HDF5Memory::open(&path).unwrap();
+    let reopened = HDF5Memory::open_read_only(&path).unwrap();
    for sess in 0..5 {
        let summary = reopened
            .get_session_summary(&format!("sess_{sess}"))
@@ -460,7 +460,7 @@ fn test_snapshot_and_continue() {
    assert_eq!(snap_mem.count(), 50);

    // Original should have 100
-    let orig_mem = HDF5Memory::open(&path).unwrap();
+    let orig_mem = HDF5Memory::open_read_only(&path).unwrap();
    assert_eq!(orig_mem.count(), 100);
 }

@@ -483,7 +483,7 @@ fn test_config_persistence_across_ops() {
    mem.add_session("s1", 0, 0, "ch", "summary").unwrap();
    mem.add_entity("Entity", "type", -1).unwrap();

-    let reopened = HDF5Memory::open(&path).unwrap();
+    let reopened = HDF5Memory::open_read_only(&path).unwrap();
    assert_eq!(reopened.config().embedding_dim, 128);
    assert_eq!(reopened.config().embedder, "custom:my-embedder-v2");
    assert_eq!(reopened.config().chunk_size, 2048);
@@ -695,7 +695,7 @@ fn test_large_text_chunks() {
    mem.save_batch(entries).unwrap();

    // Reopen and verify
-    let reopened = HDF5Memory::open(&path).unwrap();
+    let reopened = HDF5Memory::open_read_only(&path).unwrap();
    assert_eq!(reopened.count(), 10);

    let (_, cache, _, _) = read_cache(&path);
@@ -752,7 +752,7 @@ fn test_interleaved_sessions_entries() {
    mem.flush_wal().unwrap();

    // Verify
-    let reopened = HDF5Memory::open(&path).unwrap();
+    let reopened = HDF5Memory::open_read_only(&path).unwrap();
    assert_eq!(reopened.count(), 6);
    assert_eq!(
        reopened.get_session_summary("s1").unwrap().as_deref(),
@@ -806,7 +806,7 @@ fn test_knowledge_graph_with_embeddings() {
    mem.add_relation(e_python, e_hdf5, "reads", 0.9).unwrap();

    // Verify entity-embedding linkage persists
-    let reopened = HDF5Memory::open(&path).unwrap();
+    let reopened = HDF5Memory::open_read_only(&path).unwrap();
    let rust_entity = reopened.knowledge().get_entity(e_rust).unwrap();
    assert_eq!(rust_entity.embedding_idx, idx0 as i64);

@@ -1048,7 +1048,7 @@ fn test_gpu_l2_fallback_works() {
    let tombstones = vec![0u8; 3];

    let gpu = clawhdf5_agent::gpu_search::GpuSearchBackend::try_init(&vectors, &norms, 2, 1);
-    let results = gpu.search_l2(&vec![0.0, 0.0], &vectors, &tombstones, 3);
+    let results = gpu.search_l2(&[0.0, 0.0], &vectors, &tombstones, 3);

    assert_eq!(results.len(), 3);
    assert_eq!(results[0].0, 0);
@@ -1099,7 +1099,7 @@ fn test_mmap_reader_direct_access() {

    // Open via MmapReader directly
    let mmap = clawhdf5_io::MmapReader::open(&path).unwrap();
-    assert!(mmap.len() > 0);
+    assert!(!mmap.is_empty());
    // Verify we can read bytes at specific offsets
    let bytes = mmap.read_at(0, 8);
    assert!(bytes.is_some());
@@ -1144,9 +1144,11 @@ fn test_strategy_reports_backend() {
    let tombstones = vec![0u8; n];
    let query = vectors[0].clone();

+    let flat: Vec<f32> = vectors.iter().flatten().copied().collect();
    let (_, metrics) = strategy::search_with_metrics(
        &query,
        &vectors,
+        &flat,
        &norms,
        &tombstones,
        5,
@@ -165,3 +165,72 @@ 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);
+}
@@ -137,10 +137,10 @@ fn bench_hit_at_1_1014_records() {
            0.3,
            1,
        );
-        if let Some((top_idx, _)) = results.first() {
-            if *top_idx == target_indices[qi] {
-                hits += 1;
-            }
+        if let Some((top_idx, _)) = results.first()
+            && *top_idx == target_indices[qi]
+        {
+            hits += 1;
        }
    }

@@ -105,7 +105,7 @@ fn test_heavy_tombstoning() {
    assert_eq!(mem.count_active(), 5000);

    // Verify persistence
-    let reopened = HDF5Memory::open(&path).unwrap();
+    let reopened = HDF5Memory::open_read_only(&path).unwrap();
    assert_eq!(reopened.count(), 5000);
 }

@@ -163,7 +163,7 @@ fn test_large_embeddings_1536() {
    assert_eq!(mem.count(), 10_000);

    // Verify persistence
-    let reopened = HDF5Memory::open(&path).unwrap();
+    let reopened = HDF5Memory::open_read_only(&path).unwrap();
    assert_eq!(reopened.count(), 10_000);

    // Verify search works on large dims
@@ -545,7 +545,7 @@ fn test_delete_all_entries() {
    assert_eq!(mem.count(), 0);

    // Verify persistence
-    let reopened = HDF5Memory::open(&path).unwrap();
+    let reopened = HDF5Memory::open_read_only(&path).unwrap();
    assert_eq!(reopened.count(), 0);
 }

@@ -639,7 +639,7 @@ fn test_unicode_content() {
    ];
    mem.save_batch(entries).unwrap();

-    let reopened = HDF5Memory::open(&path).unwrap();
+    let reopened = HDF5Memory::open_read_only(&path).unwrap();
    assert_eq!(reopened.count(), 3);

    let (_, cache, _, _) = clawhdf5_agent::storage::read_from_disk(&path).unwrap();
@@ -685,6 +685,6 @@ fn test_rapid_save_delete_cycles() {
    assert_eq!(removed, 250);
    assert_eq!(mem.count(), 250);

-    let reopened = HDF5Memory::open(&path).unwrap();
+    let reopened = HDF5Memory::open_read_only(&path).unwrap();
    assert_eq!(reopened.count(), 250);
 }
@@ -0,0 +1,213 @@
+//! Property tests for the write-ahead log.
+//!
+//! A deterministic generator (no external crates, reproducible from the seed
+//! printed on failure) drives thousands of cases through two properties:
+//!
+//! 1. **Round trip** — whatever was appended is read back, in order, intact.
+//! 2. **Prefix under corruption** — after *any* damage to the file (bit flips,
+//!    truncation, inserted or deleted bytes, duplicated or reordered regions),
+//!    reading never panics and yields an exact *prefix* of what was written.
+//!    This is the guarantee the chained CRC exists to provide: replay may stop
+//!    early, but it never returns a corrupted, reordered, or invented entry.
+
+use clawhdf5_agent::wal::{WalEntry, WalEntryType, WalFile};
+
+/// SplitMix64: tiny, well-distributed, and fully determined by its seed.
+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 below(&mut self, n: usize) -> usize {
+        (self.next() % n.max(1) as u64) as usize
+    }
+
+    fn string(&mut self, max_len: usize) -> String {
+        const ALPHABET: &[char] = &['a', 'Z', '0', ' ', '\n', '\0', 'é', '漢', '🦀', '"'];
+        (0..self.below(max_len + 1))
+            .map(|_| ALPHABET[self.below(ALPHABET.len())])
+            .collect()
+    }
+}
+
+/// What a test appended, in a form comparable with what is read back.
+#[derive(Debug, Clone, PartialEq)]
+enum Logged {
+    Save(String, Vec<u32>, String, String, String, u64),
+    Update(usize, String, Vec<u32>, u64),
+    Tombstone(usize, u64),
+}
+
+fn logged(entry: &WalEntry) -> Logged {
+    // Compare floats by bit pattern so NaN payloads and -0.0 count as intact.
+    let bits: Vec<u32> = entry.embedding.iter().map(|f| f.to_bits()).collect();
+    let ts = entry.timestamp.to_bits();
+    match entry.entry_type {
+        WalEntryType::Save => Logged::Save(
+            entry.chunk.clone(),
+            bits,
+            entry.source_channel.clone(),
+            entry.session_id.clone(),
+            entry.tags.clone(),
+            ts,
+        ),
+        WalEntryType::Update => {
+            Logged::Update(entry.update_index.unwrap(), entry.chunk.clone(), bits, ts)
+        }
+        WalEntryType::Tombstone => Logged::Tombstone(entry.tombstone_index.unwrap(), ts),
+        WalEntryType::ActivationUpdate => unreachable!("never written by these tests"),
+    }
+}
+
+/// Append a random mix of records; return what was written.
+fn write_random_wal(path: &std::path::Path, rng: &mut Rng) -> Vec<Logged> {
+    let mut wal = WalFile::open(path).unwrap();
+    let mut written = Vec::new();
+    for _ in 0..rng.below(12) {
+        let timestamp = f64::from_bits(rng.next());
+        if rng.below(5) == 0 {
+            let index = rng.below(1000);
+            wal.append_tombstone(index, timestamp).unwrap();
+            written.push(Logged::Tombstone(index, timestamp.to_bits()));
+            continue;
+        }
+        let update_index = (rng.below(4) == 0).then(|| rng.below(1000));
+        let entry = WalEntry {
+            entry_type: if update_index.is_some() {
+                WalEntryType::Update
+            } else {
+                WalEntryType::Save
+            },
+            timestamp,
+            chunk: rng.string(40),
+            embedding: (0..rng.below(9))
+                .map(|_| f32::from_bits(rng.next() as u32))
+                .collect(),
+            source_channel: rng.string(8),
+            session_id: rng.string(8),
+            tags: rng.string(8),
+            tombstone_index: None,
+            update_index,
+        };
+        wal.append_save(&entry).unwrap();
+        written.push(logged(&entry));
+    }
+    written
+}
+
+fn read_back(path: &std::path::Path) -> Option<Vec<Logged>> {
+    WalFile::read_entries(path)
+        .ok()
+        .map(|entries| entries.iter().map(logged).collect())
+}
+
+#[test]
+fn everything_appended_is_read_back_intact() {
+    let dir = tempfile::TempDir::new().unwrap();
+    for seed in 0..300u64 {
+        let path = dir.path().join(format!("rt-{seed}.wal"));
+        let written = write_random_wal(&path, &mut Rng(seed));
+        assert_eq!(read_back(&path).unwrap(), written, "seed {seed}");
+        // Reopening (which scans and repositions) must not disturb anything.
+        drop(WalFile::open(&path).unwrap());
+        assert_eq!(
+            read_back(&path).unwrap(),
+            written,
+            "seed {seed} after reopen"
+        );
+    }
+}
+
+/// Damage `bytes` in one of several ways.
+fn corrupt(bytes: &mut Vec<u8>, rng: &mut Rng) {
+    if bytes.is_empty() {
+        return;
+    }
+    match rng.below(7) {
+        0 => {
+            let i = rng.below(bytes.len());
+            bytes[i] ^= 1 << rng.below(8);
+        }
+        1 => bytes.truncate(rng.below(bytes.len())),
+        2 => {
+            let i = rng.below(bytes.len() + 1);
+            bytes.insert(i, rng.next() as u8);
+        }
+        3 => {
+            let i = rng.below(bytes.len());
+            bytes.remove(i);
+        }
+        4 => {
+            // Duplicate a region in place (a replayed/duplicated entry).
+            let a = rng.below(bytes.len());
+            let b = a + rng.below(bytes.len() - a);
+            let region = bytes[a..b].to_vec();
+            let at = rng.below(bytes.len() + 1);
+            bytes.splice(at..at, region);
+        }
+        5 => {
+            // Swap two regions (reordered entries).
+            let mid = rng.below(bytes.len());
+            bytes.rotate_left(mid);
+        }
+        _ => {
+            let i = rng.below(bytes.len());
+            let n = rng.below(bytes.len() - i + 1);
+            for b in &mut bytes[i..i + n] {
+                *b = rng.next() as u8;
+            }
+        }
+    }
+}
+
+#[test]
+fn any_corruption_yields_a_prefix_never_a_wrong_entry() {
+    let dir = tempfile::TempDir::new().unwrap();
+    let mut shortened = 0u32;
+    for seed in 0..1500u64 {
+        let mut rng = Rng(seed ^ 0xC0FF_EE00);
+        let path = dir.path().join("c.wal");
+        let _ = std::fs::remove_file(&path);
+        let written = write_random_wal(&path, &mut rng);
+
+        let mut bytes = std::fs::read(&path).unwrap();
+        for _ in 0..=rng.below(3) {
+            corrupt(&mut bytes, &mut rng);
+        }
+        std::fs::write(&path, &bytes).unwrap();
+
+        // An unreadable header is a clean error; anything else is a prefix.
+        if let Some(read) = read_back(&path) {
+            assert!(
+                read.len() <= written.len() && read[..] == written[..read.len()],
+                "seed {seed}: read {read:?}\nis not a prefix of {written:?}"
+            );
+            if read.len() < written.len() {
+                shortened += 1;
+            }
+            // Opening for append repairs the tail; what was readable stays so,
+            // and a new entry lands right after it.
+            if let Ok(mut wal) = WalFile::open(&path) {
+                wal.append_tombstone(7, 1.0).unwrap();
+                drop(wal);
+                let mut expected = read.clone();
+                expected.push(Logged::Tombstone(7, 1.0f64.to_bits()));
+                assert_eq!(
+                    read_back(&path).unwrap(),
+                    expected,
+                    "seed {seed} after repair"
+                );
+            }
+        }
+    }
+    assert!(
+        shortened > 100,
+        "corruption rarely took effect: {shortened}"
+    );
+}
@@ -1,6 +1,6 @@
 [package]
 name = "clawhdf5-android"
-version = "2.1.0"
+version = "2.6.0"
 edition = "2024"
 description = "Android JNI bridge for edgehdf5-memory HDF5 backend"
 license = "MIT"
@@ -1,17 +1,18 @@
 [package]
 name = "clawhdf5-ann"
-version = "2.1.0"
+version = "2.6.0"
 edition = "2024"
 description = "HNSW approximate nearest neighbor index stored as HDF5"
 license = "MIT"
-repository = "https://github.com/redclawsystems/clawhdf5"
+repository = "https://git.redclaw.dev/quantumclaw/clawhdf5"
 readme = "README.md"
 keywords = ["hdf5", "ann", "hnsw", "nearest-neighbor"]
 categories = ["algorithms", "science"]

 [dependencies]
-clawhdf5-format = { path = "../clawhdf5-format", version = "2.1.0" }
-clawhdf5-io = { path = "../clawhdf5-io", version = "2.1.0" }
+clawhdf5-format = { path = "../clawhdf5-format", version = "2.6.0" }
+clawhdf5-io = { path = "../clawhdf5-io", version = "2.6.0" }
+clawhdf5-accel = { path = "../clawhdf5-accel", version = "2.6.0" }
 rayon = { version = "1", optional = true }

 [features]
@@ -5,4 +5,4 @@

 mod hnsw;

-pub use hnsw::{DistanceMetric, HnswIndex};
+pub use hnsw::{DistanceMetric, HnswIndex, Storage};
@@ -1,6 +1,6 @@
 [package]
 name = "clawhdf5-bench"
-version = "2.1.0"
+version = "2.6.0"
 edition = "2024"
 description = "Benchmark harnesses for clawhdf5-agent (Track 8)"
 license = "MIT"
@@ -13,6 +13,14 @@ path = "src/bin/longmemeval_bench.rs"
 name = "memory_arena"
 path = "src/bin/memory_arena.rs"

+[[bin]]
+name = "read_harness"
+path = "src/bin/read_harness.rs"
+
+[[bin]]
+name = "search_harness"
+path = "src/bin/search_harness.rs"
+
 [[bin]]
 name = "footprint_bench"
 path = "src/bin/footprint_bench.rs"
@@ -48,6 +56,9 @@ harness = false

 [dependencies]
 clawhdf5-agent = { path = "../clawhdf5-agent" }
+clawhdf5-ann = { path = "../clawhdf5-ann" }
+clawhdf5 = { path = "../clawhdf5" }
+clawhdf5-format = { path = "../clawhdf5-format" }
 clawhdf5-io = { path = "../clawhdf5-io" }
 mpi = { version = "0.8", optional = true }
 serde = { workspace = true }
@@ -22,7 +22,9 @@
 use std::time::Instant;

 use clawhdf5_agent::bm25::BM25Index;
-use clawhdf5_agent::consolidation::{ConsolidationConfig, ConsolidationEngine, MemorySource};
+use clawhdf5_agent::consolidation::{
+    ConsolidationConfig, ConsolidationEngine, TrustedSource, UntrustedSource,
+};
 use clawhdf5_agent::hybrid::hybrid_search;

 const EMBEDDING_DIM: usize = 384;
@@ -232,7 +234,7 @@ fn run_quality_benchmark() {
    for i in 0..SIGNAL_KEYWORDS.len() {
        let chunk = make_signal_content(i);
        let embedding = make_embedding(i * 1000);
-        let id = engine.add_memory(chunk, embedding, MemorySource::Correction, now);
+        let id = engine.add_trusted_memory(chunk, embedding, TrustedSource::Correction, now);
        signal_ids.push(id);
    }

@@ -240,7 +242,12 @@ fn run_quality_benchmark() {
    for i in 0..990 {
        let chunk = make_noise_content(i);
        let embedding = make_embedding(i + 100);
-        engine.add_memory(chunk, embedding, MemorySource::System, now + i as f64 * 0.1);
+        engine.add_trusted_memory(
+            chunk,
+            embedding,
+            TrustedSource::System,
+            now + i as f64 * 0.1,
+        );
    }

    println!("  → Inserted {} records total", engine.records().len());
@@ -333,7 +340,7 @@ fn run_cycle_time_benchmark() {
        for i in 0..n {
            let chunk = make_noise_content(i);
            let embedding = make_embedding(i);
-            engine.add_memory(chunk, embedding, MemorySource::User, now + i as f64);
+            engine.add_memory(chunk, embedding, UntrustedSource::User, now + i as f64);
        }

        // Warmup
@@ -344,7 +351,7 @@ fn run_cycle_time_benchmark() {
        for i in n..(n * 2) {
            let chunk = make_noise_content(i);
            let embedding = make_embedding(i);
-            engine.add_memory(chunk, embedding, MemorySource::User, now + i as f64);
+            engine.add_memory(chunk, embedding, UntrustedSource::User, now + i as f64);
        }

        // Timed consolidation
@@ -410,13 +417,13 @@ fn run_memory_reduction_benchmark() {
        for i in 0..signal_count {
            let chunk = make_signal_content(i % SIGNAL_KEYWORDS.len());
            let emb = make_embedding(i * 999);
-            let id = engine.add_memory(chunk, emb, MemorySource::Correction, now);
+            let id = engine.add_trusted_memory(chunk, emb, TrustedSource::Correction, now);
            signal_ids.push(id);
        }
        for i in 0..noise_count {
            let chunk = make_noise_content(i);
            let emb = make_embedding(i + 200);
-            engine.add_memory(chunk, emb, MemorySource::System, now + i as f64 * 0.1);
+            engine.add_trusted_memory(chunk, emb, TrustedSource::System, now + i as f64 * 0.1);
        }

        // Access signal records heavily
@@ -55,44 +55,150 @@ use std::time::{Duration, Instant};
 #[path = "longmemeval_bench/embedder.rs"]
 mod embedder;

-use clawhdf5_agent::{AgentMemory, HDF5Memory, MemoryConfig, MemoryEntry};
+use clawhdf5_agent::bm25::TokenFilter;
+use clawhdf5_agent::hybrid::Fusion;
+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;

+/// 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}"),
+    };
+    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
+        ),
+    }
+}
+
 /// A retrieval configuration: how much of the score comes from each stage.
 #[derive(Clone, Copy)]
 struct Mode {
    label: &'static str,
-    vector_weight: f32,
-    keyword_weight: f32,
+    /// How the two retrieval stages are combined into one ranking.
+    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 {
+    const fn weighted(label: &'static str, vector: f32, keyword: f32) -> Self {
+        Self {
+            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;
+        self
+    }
 }

 /// The only mode available without real embeddings. Passing zero vectors with
 /// `vector_weight = 0.0` is what made the vector stage inert.
-const BM25_ONLY: Mode = Mode {
-    label: "BM25 only (vector stage inert)",
-    vector_weight: 0.0,
-    keyword_weight: 1.0,
-};
+const BM25_ONLY: Mode = Mode::weighted("BM25 only (vector stage inert)", 0.0, 1.0);
 #[cfg(feature = "embeddings")]
-const VECTOR_ONLY: Mode = Mode {
-    label: "Vector only (MiniLM + HNSW)",
-    vector_weight: 1.0,
-    keyword_weight: 0.0,
-};
+const VECTOR_ONLY: Mode = Mode::weighted("Vector only (MiniLM + HNSW)", 1.0, 0.0);
 /// Tuned by `--sweep` over the full haystack. The former 0.7/0.3 was a
 /// documented default that had never been searched, and the sweep found it
 /// strictly dominated: 0.4/0.6 is better on Hit@1, Hit@5, Hit@10 and MRR at
 /// both granularities.
 #[cfg(feature = "embeddings")]
-const HYBRID: Mode = Mode {
-    label: "Hybrid (0.4 vector / 0.6 BM25, tuned)",
-    vector_weight: 0.4,
-    keyword_weight: 0.6,
+const HYBRID: Mode = Mode::weighted("Hybrid (0.4 vector / 0.6 BM25, tuned)", 0.4, 0.6);
+
+/// Reciprocal rank fusion, the documented alternative to the weighted sum.
+/// It ignores score magnitudes, so there is nothing to tune — which is the
+/// claim being tested.
+#[cfg(feature = "embeddings")]
+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");
+
 /// Every 0.1 step of vector weight, keyword weight taking the remainder.
 ///
 /// Labels are leaked to `&'static str` because `Mode::label` is a `&'static
@@ -104,11 +210,11 @@ fn sweep_modes() -> Vec<Mode> {
    (0..=10)
        .map(|i| {
            let v = i as f32 / 10.0;
-            Mode {
-                label: Box::leak(format!("sweep v={v:.1} / k={:.1}", 1.0 - v).into_boxed_str()),
-                vector_weight: v,
-                keyword_weight: 1.0 - v,
-            }
+            Mode::weighted(
+                Box::leak(format!("sweep v={v:.1} / k={:.1}", 1.0 - v).into_boxed_str()),
+                v,
+                1.0 - v,
+            )
        })
        .collect()
 }
@@ -181,6 +287,37 @@ 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)
 }

 // ---------------------------------------------------------------------------
@@ -199,11 +336,21 @@ 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
    }
@@ -261,6 +408,16 @@ 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,
 }

@@ -276,19 +433,26 @@ fn evaluate_question(
    config.compact_threshold = 0.0;

    let mut memory = HDF5Memory::create(config).expect("failed to create HDF5Memory");
+    memory.set_token_filter(mode.tokens);

    // 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),
@@ -302,7 +466,6 @@ fn evaluate_question(
                },
            });
            turn_has_answer.push(turn.has_answer);
-            ts += 1.0;
        }
    }

@@ -319,17 +482,87 @@ 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(
-        &query_emb,
-        &q.question,
-        mode.vector_weight,
-        mode.keyword_weight,
-        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;
@@ -384,6 +617,7 @@ fn evaluate_question(
        hit5_turn,
        hit10_turn,
        rr_turn,
+        newest_gold_first,
        latency,
    }
 }
@@ -472,10 +706,7 @@ fn print_report(
    println!("  LongMemEval Benchmark  —  {}", mode.label);
    println!("=================================================================");
    println!();
-    println!(
-        "Mode: vector_weight={:.1} / keyword_weight={:.1}",
-        mode.vector_weight, mode.keyword_weight
-    );
+    println!("Mode: {}", describe(mode));
    println!();
    println!("Scoring target: RETRIEVAL RECALL (did the gold memory land in top-k).");
    println!("  No answer is generated or scored. This is NOT the official");
@@ -538,6 +769,24 @@ 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!(
@@ -602,10 +851,7 @@ fn print_report(
    println!("```json");
    println!("{{");
    println!("  \"benchmark\": \"longmemeval\",");
-    println!(
-        "  \"mode\": \"vector_{:.1}_keyword_{:.1}\",",
-        mode.vector_weight, mode.keyword_weight
-    );
+    println!("  \"mode\": \"{}\",", describe(mode));
    println!("  \"dataset_variant\": \"{}\",", profile.variant());
    println!("  \"scoring_target\": \"retrieval_recall\",");
    println!("  \"k\": 10,");
@@ -654,6 +900,14 @@ 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}",
@@ -676,6 +930,8 @@ 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() {
@@ -684,6 +940,16 @@ fn main() {
                limit = Some(v.parse().expect("--limit must be a positive integer"));
            }
            "--sweep" => sweep = true,
+            "--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"));
            }
@@ -702,6 +968,9 @@ 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\
                     --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."
@@ -767,19 +1036,34 @@ 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, VECTOR_ONLY, HYBRID]
+                vec![
+                    BM25_ONLY,
+                    VECTOR_ONLY,
+                    HYBRID,
+                    RRF,
+                    BM25_STEMMED,
+                    HYBRID_STEMMED,
+                    hybrid_rerank_metadata_only(),
+                    hybrid_rerank_blended(),
+                ]
            }
        }
        #[cfg(not(feature = "embeddings"))]
        {
-            vec![BM25_ONLY]
+            vec![BM25_ONLY, BM25_STEMMED]
        }
    } else {
        if sweep {
            eprintln!("warning: --sweep needs --embeddings; running BM25 only");
        }
-        vec![BM25_ONLY]
+        // Stemming is a property of the keyword stage, so it can be compared
+        // without a model.
+        vec![BM25_ONLY, BM25_STEMMED]
    };

    for (mode_idx, mode) in modes.iter().enumerate() {
@@ -882,6 +1166,14 @@ 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);
@@ -893,3 +1185,30 @@ 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");
+    }
+}
@@ -0,0 +1,176 @@
+//! HDF5 read-path measurement harness: full reads vs. hyperslab selections on
+//! a chunked 2-D dataset, compressed and uncompressed, plus a contiguous one.
+//!
+//! The question it answers for every read-path change: does the cost of a
+//! selection scale with the *selection*, or with the whole dataset?
+//!
+//! ```text
+//! cargo run --release -p clawhdf5-bench --bin read_harness
+//! cargo run --release -p clawhdf5-bench --bin read_harness -- --large   # 512 MB
+//! ```
+
+use std::time::{Duration, Instant};
+
+use clawhdf5::{File, FileBuilder};
+use clawhdf5_format::selection::Selection;
+
+const CHUNK: u64 = 256;
+
+struct Layout {
+    name: &'static str,
+    chunked: bool,
+    deflate: bool,
+}
+
+const LAYOUTS: [Layout; 3] = [
+    Layout {
+        name: "chunked + deflate",
+        chunked: true,
+        deflate: true,
+    },
+    Layout {
+        name: "chunked",
+        chunked: true,
+        deflate: false,
+    },
+    Layout {
+        name: "contiguous",
+        chunked: false,
+        deflate: false,
+    },
+];
+
+/// Smooth-ish, compressible data whose value encodes its position, so a read
+/// can be verified exactly.
+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) {
+    let data: Vec<f64> = (0..rows)
+        .flat_map(|r| (0..cols).map(move |c| value(r, c)))
+        .collect();
+    let mut builder = FileBuilder::new();
+    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]);
+        }
+        if layout.deflate {
+            ds.with_deflate(4);
+        }
+    }
+    builder.write(path).unwrap();
+}
+
+fn median(mut samples: Vec<Duration>) -> Duration {
+    samples.sort();
+    samples[samples.len() / 2]
+}
+
+fn time<T>(reps: usize, mut f: impl FnMut() -> T) -> Duration {
+    median(
+        (0..reps)
+            .map(|_| {
+                let t = Instant::now();
+                std::hint::black_box(f());
+                t.elapsed()
+            })
+            .collect(),
+    )
+}
+
+fn slab(start: [u64; 2], count: [u64; 2]) -> Selection {
+    Selection::Hyperslab {
+        start: start.to_vec(),
+        stride: vec![1, 1],
+        count: count.to_vec(),
+        block: vec![1, 1],
+    }
+}
+
+fn main() {
+    let large = std::env::args().any(|a| a == "--large");
+    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) {
+        eprintln!("warning: debug build — numbers are meaningless. Use --release.");
+    }
+
+    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;
+
+    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"
+    );
+
+    // (label, selection, elements selected)
+    let selections: Vec<(&str, Selection, u64)> = vec![
+        (
+            "64 x 64 window (1 chunk)",
+            slab([300, 300], [64, 64]),
+            64 * 64,
+        ),
+        (
+            "512 x 512 window (4-9 chunks)",
+            slab([1000, 700], [512, 512]),
+            512 * 512,
+        ),
+        ("one row", slab([rows / 2, 0], [1, cols]), cols),
+        ("one column", slab([0, cols / 2], [rows, 1]), rows),
+    ];
+
+    println!("| layout | read | selected | time ms | MB/s of selection | vs full read |");
+    println!("|---|---|---:|---:|---:|---:|");
+    for (i, layout) in LAYOUTS.iter().enumerate() {
+        // Fresh handle per layout so one dataset's cached chunks don't help
+        // (or evict) another's.
+        let file = File::open(&path).unwrap();
+        let ds = file.dataset(&format!("d{i}")).unwrap();
+
+        let full_cold = time(1, || ds.read_f64().unwrap());
+        let full = time(3, || ds.read_f64().unwrap());
+        println!(
+            "| {} | full (first) | {total_mb:.0} MB | {:.1} | {:.0} | |",
+            layout.name,
+            full_cold.as_secs_f64() * 1e3,
+            total_mb / full_cold.as_secs_f64()
+        );
+        println!(
+            "| {} | full (repeat) | {total_mb:.0} MB | {:.1} | {:.0} | 1.00x |",
+            layout.name,
+            full.as_secs_f64() * 1e3,
+            total_mb / full.as_secs_f64()
+        );
+
+        for (label, selection, elements) in &selections {
+            // A fresh handle again: measure the selection on its own, not
+            // served from chunks the full read just cached.
+            let file = File::open(&path).unwrap();
+            let ds = file.dataset(&format!("d{i}")).unwrap();
+            let got = ds.read_f64_selection(selection).unwrap();
+            assert_eq!(got.len() as u64, *elements, "{label}");
+            if let Selection::Hyperslab { start, .. } = selection {
+                assert_eq!(got[0], value(start[0], start[1]), "{label}: wrong data");
+            }
+            let took = time(5, || {
+                let file = File::open(&path).unwrap();
+                let ds = file.dataset(&format!("d{i}")).unwrap();
+                ds.read_f64_selection(selection).unwrap()
+            });
+            let mb = (*elements * 8) as f64 / (1 << 20) as f64;
+            println!(
+                "| {} | {label} | {:.2} MB | {:.2} | {:.0} | {:.3}x |",
+                layout.name,
+                mb,
+                took.as_secs_f64() * 1e3,
+                mb / took.as_secs_f64(),
+                took.as_secs_f64() / full_cold.as_secs_f64()
+            );
+        }
+    }
+}
@@ -0,0 +1,679 @@
+//! Search measurement harness: recall vs. speed for the HNSW index, and
+//! end-to-end `hybrid_search` latency as the store grows.
+//!
+//! Every search-path change should be justified by a before/after run of this
+//! binary. It reports, for deterministic synthetic data:
+//!
+//! * **ANN** — index build time, and for each `ef`: recall@10 against an exact
+//!   brute-force scan, queries/second, and p50/p99 latency.
+//! * **End to end** — `HDF5Memory`: ingest time, checkpoint time, `open()`
+//!   time, the one-off cold index build (first query ever), the first query
+//!   after a reopen, and steady-state `hybrid_search` p50/p99 at each size.
+//!
+//! Data is *clustered* (points = cluster centre + noise, unit-normalised), not
+//! uniform: uniform random high-dimensional vectors are nearly equidistant,
+//! which makes recall numbers meaningless and is nothing like embeddings.
+//!
+//! ```text
+//! cargo run --release -p clawhdf5-bench --bin search_harness            # 1K, 10K
+//! 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
+//! ```
+
+use std::time::{Duration, Instant};
+
+use clawhdf5_agent::{AgentMemory, HDF5Memory, MemoryConfig, MemoryEntry};
+use clawhdf5_ann::{DistanceMetric, HnswIndex, Storage};
+
+const DIM: usize = 384;
+const K: usize = 10;
+const N_QUERIES: usize = 200;
+const HNSW_M: usize = 16;
+const HNSW_EF_CONSTRUCTION: usize = 64;
+const EF_VALUES: [usize; 5] = [16, 32, 64, 128, 256];
+
+// ---------------------------------------------------------------------------
+// Deterministic data
+// ---------------------------------------------------------------------------
+
+struct Rng(u64);
+
+impl Rng {
+    fn next_u64(&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)
+    }
+
+    /// Uniform in [0, 1).
+    fn unit(&mut self) -> f32 {
+        (self.next_u64() >> 40) as f32 / (1u64 << 24) as f32
+    }
+
+    /// Approximately standard normal (sum of uniforms).
+    fn gauss(&mut self) -> f32 {
+        let sum: f32 = (0..6).map(|_| self.unit()).sum();
+        (sum - 3.0) * std::f32::consts::SQRT_2
+    }
+
+    fn below(&mut self, n: usize) -> usize {
+        (self.next_u64() % n as u64) as usize
+    }
+}
+
+fn normalize(v: &mut [f32]) {
+    let norm = v.iter().map(|x| x * x).sum::<f32>().sqrt();
+    if norm > 0.0 {
+        v.iter_mut().for_each(|x| *x /= norm);
+    }
+}
+
+struct Dataset {
+    vectors: Vec<Vec<f32>>,
+    queries: Vec<Vec<f32>>,
+    /// Cluster id of each vector (used to give records topical text).
+    cluster_of: Vec<usize>,
+    query_cluster: Vec<usize>,
+}
+
+/// `--uniform`: isotropic random unit vectors instead of clusters. Not a
+/// realistic workload, but a useful second distribution — a recall problem
+/// 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);
+
+/// `--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) {
+        let random_unit = |rng: &mut Rng| {
+            let mut v: Vec<f32> = (0..DIM).map(|_| rng.gauss()).collect();
+            normalize(&mut v);
+            v
+        };
+        return Dataset {
+            vectors: (0..n).map(|_| random_unit(&mut rng)).collect(),
+            queries: (0..N_QUERIES).map(|_| random_unit(&mut rng)).collect(),
+            cluster_of: vec![0; n],
+            query_cluster: vec![0; N_QUERIES],
+        };
+    }
+    let n_clusters = (n / 100).clamp(8, 512);
+    let centres: Vec<Vec<f32>> = (0..n_clusters)
+        .map(|_| {
+            let mut c: Vec<f32> = (0..DIM).map(|_| rng.gauss()).collect();
+            normalize(&mut c);
+            c
+        })
+        .collect();
+    let point = |rng: &mut Rng, cluster: usize| {
+        // Noise comparable to the centre's per-dimension magnitude, so
+        // clusters overlap and the nearest neighbours are non-trivial.
+        let scale = 0.6 / (DIM as f32).sqrt();
+        let mut v: Vec<f32> = centres[cluster]
+            .iter()
+            .map(|c| c + rng.gauss() * scale)
+            .collect();
+        normalize(&mut v);
+        v
+    };
+    let mut vectors = Vec::with_capacity(n);
+    let mut cluster_of = Vec::with_capacity(n);
+    for _ in 0..n {
+        let c = rng.below(n_clusters);
+        vectors.push(point(&mut rng, c));
+        cluster_of.push(c);
+    }
+    let mut queries = Vec::with_capacity(N_QUERIES);
+    let mut query_cluster = Vec::with_capacity(N_QUERIES);
+    for _ in 0..N_QUERIES {
+        let c = rng.below(n_clusters);
+        queries.push(point(&mut rng, c));
+        query_cluster.push(c);
+    }
+    Dataset {
+        vectors,
+        queries,
+        cluster_of,
+        query_cluster,
+    }
+}
+
+const WORDS: &[&str] = &[
+    "deploy", "latency", "cache", "schema", "index", "vector", "memory", "agent", "kernel",
+    "buffer", "socket", "thread", "tensor", "gradient", "ledger", "invoice", "meeting", "roadmap",
+    "customer", "contract", "sensor", "orbit", "protein", "genome", "harbor", "bridge", "engine",
+    "battery", "harvest", "weather", "museum", "recipe",
+];
+
+/// Text whose vocabulary is biased by cluster, so keyword and vector signals
+/// agree the way they do for real embedded text.
+fn text_for(cluster: usize, i: usize, rng: &mut Rng) -> String {
+    let topic = [
+        WORDS[cluster % WORDS.len()],
+        WORDS[(cluster / 7 + 3) % WORDS.len()],
+    ];
+    let mut words = Vec::with_capacity(14);
+    for j in 0..14 {
+        if j % 3 == 0 {
+            words.push(topic[j / 3 % 2]);
+        } else {
+            words.push(WORDS[rng.below(WORDS.len())]);
+        }
+    }
+    format!("record {i}: {}", words.join(" "))
+}
+
+// ---------------------------------------------------------------------------
+// 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
+        .iter()
+        .enumerate()
+        .map(|(i, v)| (i, v.iter().zip(query).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.truncate(k);
+    scored.into_iter().map(|(i, _)| i).collect()
+}
+
+struct Latency {
+    p50: Duration,
+    p99: Duration,
+    qps: f64,
+}
+
+fn summarize(mut samples: Vec<Duration>) -> Latency {
+    samples.sort();
+    let total: Duration = samples.iter().sum();
+    let at = |q: f64| samples[((samples.len() - 1) as f64 * q).round() as usize];
+    Latency {
+        p50: at(0.50),
+        p99: at(0.99),
+        qps: samples.len() as f64 / total.as_secs_f64(),
+    }
+}
+
+/// 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);
+
+// 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() {
+            LIVE_BYTES.fetch_add(layout.size() as i64, std::sync::atomic::Ordering::Relaxed);
+        }
+        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() {
+            LIVE_BYTES.fetch_add(
+                new_size as i64 - layout.size() as i64,
+                std::sync::atomic::Ordering::Relaxed,
+            );
+        }
+        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
+}
+
+fn mib(bytes: u64) -> f64 {
+    bytes as f64 / (1 << 20) as f64
+}
+
+fn micros(d: Duration) -> f64 {
+    d.as_secs_f64() * 1e6
+}
+
+fn millis(d: Duration) -> f64 {
+    d.as_secs_f64() * 1e3
+}
+
+// ---------------------------------------------------------------------------
+// ANN: recall vs speed
+// ---------------------------------------------------------------------------
+
+fn bench_ann(n: usize, json: &mut Vec<serde_json::Value>) {
+    let data = make_dataset(n, 0xA11CE ^ n as u64);
+    let truth: Vec<Vec<usize>> = data
+        .queries
+        .iter()
+        .map(|q| exact_top_k(&data.vectors, q, K))
+        .collect();
+
+    let started = Instant::now();
+    let index = HnswIndex::build_with(
+        &data.vectors,
+        HNSW_M,
+        HNSW_EF_CONSTRUCTION,
+        DistanceMetric::Cosine,
+        storage(),
+    );
+    let build = started.elapsed();
+
+    // Exact scan baseline, for scale.
+    let exact = summarize(
+        data.queries
+            .iter()
+            .map(|q| {
+                let t = Instant::now();
+                std::hint::black_box(exact_top_k(&data.vectors, q, K));
+                t.elapsed()
+            })
+            .collect(),
+    );
+
+    println!(
+        "\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",
+        millis(build),
+        n as f64 / build.as_secs_f64(),
+        exact.qps,
+        micros(exact.p50)
+    );
+    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 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();
+        }
+        let recall = hits as f64 / (K * data.queries.len()) as f64;
+        let lat = summarize(samples);
+        println!(
+            "| {ef} | {recall:.4} | {:.0} | {:.0} | {:.0} |",
+            lat.qps,
+            micros(lat.p50),
+            micros(lat.p99)
+        );
+        json.push(serde_json::json!({
+            "bench": "hnsw", "n": n, "ef": ef, "recall_at_10": recall,
+            "qps": lat.qps, "p50_us": micros(lat.p50), "p99_us": micros(lat.p99),
+            "build_ms": millis(build),
+        }));
+    }
+}
+
+// ---------------------------------------------------------------------------
+// End to end: HDF5Memory::hybrid_search
+// ---------------------------------------------------------------------------
+
+fn bench_end_to_end(n: usize, json: &mut Vec<serde_json::Value>) {
+    let data = make_dataset(n, 0xE2E ^ n as u64);
+    let dir = tempfile::TempDir::new().unwrap();
+    let path = dir.path().join("store.h5");
+    let mut rng = Rng(7);
+
+    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 query_texts: Vec<String> = data
+        .query_cluster
+        .iter()
+        .enumerate()
+        .map(|(i, c)| text_for(*c, i, &mut rng))
+        .collect();
+
+    let mut mem = HDF5Memory::create(MemoryConfig::new(path.clone(), "bench", DIM)).unwrap();
+    let t = Instant::now();
+    mem.save_batch(entries).unwrap();
+    let ingest = t.elapsed();
+    // The very first query builds the vector and keyword indexes from
+    // scratch. It happens once per store, not once per session: the checkpoint
+    // below saves the vector index, so a later `open()` reloads it.
+    let t = Instant::now();
+    std::hint::black_box(mem.hybrid_search(&data.queries[1], &query_texts[1], 0.7, 0.3, K));
+    let cold_build = t.elapsed();
+
+    let t = Instant::now();
+    mem.flush_wal().unwrap();
+    let checkpoint = t.elapsed();
+    drop(mem);
+
+    let t = Instant::now();
+    let mut mem = HDF5Memory::open(&path).unwrap();
+    let open = t.elapsed();
+
+    // The first query after open pays for whatever is rebuilt lazily.
+    let t = Instant::now();
+    std::hint::black_box(mem.hybrid_search(&data.queries[0], &query_texts[0], 0.7, 0.3, K));
+    let first_query = t.elapsed();
+
+    // Fewer steady-state samples at large N: each query is currently O(N).
+    let samples_wanted = if n >= 100_000 { 20 } else { N_QUERIES.min(100) };
+    let steady = summarize(
+        (0..samples_wanted)
+            .map(|i| {
+                let t = Instant::now();
+                std::hint::black_box(mem.hybrid_search(
+                    &data.queries[i % N_QUERIES],
+                    &query_texts[i % N_QUERIES],
+                    0.7,
+                    0.3,
+                    K,
+                ));
+                t.elapsed()
+            })
+            .collect(),
+    );
+
+    println!(
+        "| {n} | {:.0} | {:.0} | {:.1} | {:.1} | {:.1} | {:.2} | {:.2} | {:.1} |",
+        millis(ingest),
+        millis(cold_build),
+        millis(checkpoint),
+        millis(open),
+        millis(first_query),
+        millis(steady.p50),
+        millis(steady.p99),
+        steady.qps
+    );
+    json.push(serde_json::json!({
+        "bench": "hybrid_search", "n": n,
+        "ingest_ms": millis(ingest), "cold_index_build_ms": millis(cold_build),
+        "checkpoint_ms": millis(checkpoint),
+        "open_ms": millis(open), "first_query_ms": millis(first_query),
+        "p50_ms": millis(steady.p50), "p99_ms": millis(steady.p99), "qps": steady.qps,
+    }));
+}
+
+// ---------------------------------------------------------------------------
+// Fusion study: does capping the keyword candidate pool change the ranking?
+// ---------------------------------------------------------------------------
+
+/// `hybrid_search` min-max normalises each signal over the candidates it is
+/// given. The vector stage supplies a pool of `max(8k, 64)`; the keyword stage
+/// supplies *every* matching record, which is what now dominates query time.
+/// This compares the current fusion with one whose keyword stage is capped to
+/// a pool, reporting how often the final top-k agree and what each costs.
+fn fusion_study(n: usize) {
+    use clawhdf5_agent::bm25::BM25Index;
+    use clawhdf5_agent::hybrid::merge_vector_keyword;
+
+    let data = make_dataset(n, 0xE2E ^ n as u64);
+    let mut rng = Rng(7);
+    let texts: Vec<String> = (0..n)
+        .map(|i| text_for(data.cluster_of[i], i, &mut rng))
+        .collect();
+    let query_texts: Vec<String> = data
+        .query_cluster
+        .iter()
+        .enumerate()
+        .map(|(i, c)| text_for(*c, i, &mut rng))
+        .collect();
+    let bm25 = BM25Index::build(&texts, &vec![0u8; n]);
+    let index = HnswIndex::build_with(
+        &data.vectors,
+        HNSW_M,
+        HNSW_EF_CONSTRUCTION,
+        DistanceMetric::Cosine,
+        storage(),
+    );
+
+    let vec_pool = (K * 8).max(64);
+    println!("\n### Fusion study, N = {n} (k = {K}, weights 0.7 / 0.3, vector pool {vec_pool})\n");
+    println!(
+        "| keyword pool | top-{K} overlap vs full | identical top-{K} | same #1 | keyword+merge µs |"
+    );
+    println!("|---:|---:|---:|---:|---:|");
+
+    let fuse = |q: usize, kw_pool: usize| -> (Vec<usize>, Duration) {
+        let vec_scores: Vec<(usize, f32)> = index
+            .search(&data.queries[q], vec_pool, vec_pool)
+            .into_iter()
+            .map(|(id, d)| (id, 1.0 - d))
+            .collect();
+        let t = Instant::now();
+        let kw = bm25.search(&query_texts[q], kw_pool);
+        let merged = merge_vector_keyword(vec_scores, kw, 0.7, 0.3, K);
+        let took = t.elapsed();
+        (merged.into_iter().map(|(id, _)| id).collect(), took)
+    };
+
+    let full: Vec<(Vec<usize>, Duration)> = (0..N_QUERIES).map(|q| fuse(q, n)).collect();
+    let full_time: Duration = full.iter().map(|f| f.1).sum();
+    println!(
+        "| all ({n}) | 1.0000 | 100.0% | 100.0% | {:.0} |",
+        micros(full_time) / N_QUERIES as f64
+    );
+    for pool in [vec_pool, vec_pool * 4, 1000] {
+        if pool >= n {
+            continue;
+        }
+        let (mut overlap, mut identical, mut same_first) = (0usize, 0usize, 0usize);
+        let mut time = Duration::ZERO;
+        for (q, (want, _)) in full.iter().enumerate() {
+            let (got, took) = fuse(q, pool);
+            time += took;
+            overlap += got.iter().filter(|id| want.contains(id)).count();
+            identical += usize::from(&got == want);
+            same_first += usize::from(got.first() == want.first());
+        }
+        println!(
+            "| {pool} | {:.4} | {:.1}% | {:.1}% | {:.0} |",
+            overlap as f64 / (K * N_QUERIES) as f64,
+            100.0 * identical as f64 / N_QUERIES as f64,
+            100.0 * same_first as f64 / N_QUERIES as f64,
+            micros(time) / N_QUERIES as f64
+        );
+    }
+}
+
+/// 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();
+    let reopened = HDF5Memory::open(&path).unwrap();
+    let after_open = heap_bytes();
+    let loaded = after_open.saturating_sub(before_open);
+    drop(reopened);
+
+    let raw = (n * DIM * 4) as u64;
+    println!(
+        "| {n} | {:.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),
+        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");
+    let ann_only = args.iter().any(|a| a == "--ann-only");
+    if args.iter().any(|a| a == "--fusion-study") {
+        for &n in if full {
+            &[10_000, 100_000][..]
+        } else {
+            &[10_000][..]
+        } {
+            fusion_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)");
+    }
+    let json_path = args
+        .iter()
+        .position(|a| a == "--json")
+        .and_then(|i| args.get(i + 1))
+        .cloned();
+    let sizes: &[usize] = if full {
+        &[1_000, 10_000, 100_000]
+    } else {
+        &[1_000, 10_000]
+    };
+
+    if cfg!(debug_assertions) {
+        eprintln!("warning: debug build — numbers are meaningless. Use --release.");
+    }
+
+    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 | 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") {
+        for &n in sizes {
+            bench_ann(n, &mut json);
+        }
+    }
+
+    if ann_only {
+        return;
+    }
+    println!("\n### End to end: `HDF5Memory::hybrid_search` (k = {K}, weights 0.7 / 0.3)\n");
+    println!(
+        "| N | ingest ms | cold index build ms | checkpoint ms | open ms | first query after open ms | p50 ms | p99 ms | QPS |"
+    );
+    println!("|---:|---:|---:|---:|---:|---:|---:|---:|---:|");
+    for &n in sizes {
+        bench_end_to_end(n, &mut json);
+    }
+
+    if let Some(path) = json_path {
+        std::fs::write(&path, serde_json::to_string_pretty(&json).unwrap()).unwrap();
+        eprintln!("wrote {path}");
+    }
+}
@@ -1,10 +1,10 @@
 [package]
 name = "clawhdf5-cli"
-version = "2.1.0"
+version = "2.6.0"
 edition = "2024"
 license = "MIT"
 description = "CLI for clawhdf5 agent memory — create, save, search, recall, stats"
-repository = "https://github.com/redclawsystems/clawhdf5"
+repository = "https://git.redclaw.dev/quantumclaw/clawhdf5"
 keywords = ["hdf5", "ai", "memory", "agent", "cli"]
 categories = ["command-line-utilities", "science"]
 readme = "../../README.md"
@@ -14,7 +14,7 @@ name = "clawhdf5"
 path = "src/main.rs"

 [dependencies]
-clawhdf5-agent = { path = "../clawhdf5-agent", version = "2.1.0" }
+clawhdf5-agent = { path = "../clawhdf5-agent", version = "2.6.0" }
 clap = { version = "4", features = ["derive", "env"] }
 serde_json = "1"
 serde = { workspace = true }
@@ -28,6 +28,10 @@ enum Commands {
        /// Enable write-ahead log
        #[arg(long)]
        wal: bool,
+        /// Store the vector index's copy of the embeddings as int8, roughly
+        /// halving a loaded store's memory at about 13% fewer queries/second
+        #[arg(long)]
+        quantized_index: bool,
    },
    /// Save a memory entry (reads JSON from stdin or --json)
    Save {
@@ -88,9 +92,15 @@ fn main() {

 fn run(cli: Cli) -> Result<(), Box<dyn std::error::Error>> {
    match cli.command {
-        Commands::Create { agent_id, dim, wal } => {
+        Commands::Create {
+            agent_id,
+            dim,
+            wal,
+            quantized_index,
+        } => {
            let mut config = MemoryConfig::new(cli.path.clone(), &agent_id, dim);
            config.wal_enabled = wal;
+            config.quantized_index = quantized_index;
            let mem = HDF5Memory::create(config)?;
            let j = serde_json::json!({
                "status": "created",
@@ -98,6 +108,7 @@ fn run(cli: Cli) -> Result<(), Box<dyn std::error::Error>> {
                "agent_id": agent_id,
                "embedding_dim": dim,
                "wal_enabled": wal,
+                "quantized_index": quantized_index,
                "count": mem.count(),
            });
            println!("{}", serde_json::to_string_pretty(&j)?);
@@ -146,7 +157,7 @@ fn run(cli: Cli) -> Result<(), Box<dyn std::error::Error>> {
        }

        Commands::Recall { index } => {
-            let mem = HDF5Memory::open(&cli.path)?;
+            let mem = HDF5Memory::open_read_only(&cli.path)?;
            match mem.get_chunk(index) {
                Some(content) => {
                    let j = serde_json::json!({ "index": index, "chunk": content });
@@ -160,7 +171,7 @@ fn run(cli: Cli) -> Result<(), Box<dyn std::error::Error>> {
        }

        Commands::Stats => {
-            let mem = HDF5Memory::open(&cli.path)?;
+            let mem = HDF5Memory::open_read_only(&cli.path)?;
            let cfg = mem.config();
            let j = serde_json::json!({
                "path": cli.path.display().to_string(),
@@ -187,7 +198,7 @@ fn run(cli: Cli) -> Result<(), Box<dyn std::error::Error>> {
        }

        Commands::AgentsMd { output } => {
-            let mem = HDF5Memory::open(&cli.path)?;
+            let mem = HDF5Memory::open_read_only(&cli.path)?;
            let md = mem.generate_agents_md();
            match output {
                Some(p) => {
@@ -199,7 +210,7 @@ fn run(cli: Cli) -> Result<(), Box<dyn std::error::Error>> {
        }

        Commands::Export => {
-            let mem = HDF5Memory::open(&cli.path)?;
+            let mem = HDF5Memory::open_read_only(&cli.path)?;
            for i in 0..mem.count() {
                if let Some(chunk) = mem.get_chunk(i) {
                    let j = serde_json::json!({ "index": i, "chunk": chunk });
@@ -1,10 +1,10 @@
 [package]
 name = "clawhdf5-derive"
-version = "2.1.0"
+version = "2.6.0"
 edition = "2024"
 description = "Derive macros for rustyhdf5 HDF5 traits"
 license = "MIT"
-repository = "https://github.com/redclawsystems/clawhdf5"
+repository = "https://git.redclaw.dev/quantumclaw/clawhdf5"
 readme = "README.md"
 keywords = ["hdf5", "derive", "macros", "science"]
 categories = ["development-tools::procedural-macro-helpers"]
@@ -1,10 +1,10 @@
 [package]
 name = "clawhdf5-filters"
-version = "2.1.0"
+version = "2.6.0"
 edition = "2024"
 description = "Filter and compression pipeline for clawhdf5"
 license = "MIT"
-repository = "https://github.com/redclawsystems/clawhdf5"
+repository = "https://git.redclaw.dev/quantumclaw/clawhdf5"
 readme = "README.md"
 keywords = ["hdf5", "compression", "deflate", "filters"]
 categories = ["compression", "science"]
@@ -1,10 +1,10 @@
 [package]
 name = "clawhdf5-format"
-version = "2.1.0"
+version = "2.6.0"
 edition = "2024"
 description = "Pure-Rust HDF5 binary format parsing and writing — no C dependencies"
 license = "MIT"
-repository = "https://github.com/redclawsystems/clawhdf5"
+repository = "https://git.redclaw.dev/quantumclaw/clawhdf5"
 readme = "README.md"
 keywords = ["hdf5", "science", "data", "binary", "no-std"]
 categories = ["parser-implementations", "science", "encoding", "no-std"]
@@ -25,7 +25,7 @@ pco = { version = "1.0", optional = true }
 [dev-dependencies]
 serde_json = "1"
 criterion = { workspace = true }
-clawhdf5-derive = { path = "../clawhdf5-derive", version = "2.1.0" }
+clawhdf5-derive = { path = "../clawhdf5-derive", version = "2.6.0" }

 [[bench]]
 name = "bench"
@@ -1,7 +1,9 @@
 //! HDF5 Attribute message parsing (message type 0x000C).

 #[cfg(not(feature = "std"))]
-use alloc::{string::String, vec::Vec};
+use alloc::{borrow::Cow, string::String, vec::Vec};
+#[cfg(feature = "std")]
+use std::borrow::Cow;

 use crate::attribute_info::AttributeInfoMessage;
 use crate::btree_v2::{BTreeV2Header, collect_btree_v2_records};
@@ -48,17 +50,64 @@ 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)
+    }
+
+    /// [`AttributeMessage::parse`] with access to the rest of the file, which
+    /// is needed when the attribute's datatype or dataspace is *shared* (v2/v3
+    /// flag bits 0/1) — e.g. an attribute created with a committed datatype.
+    /// In that case the embedded bytes are a reference to the real message,
+    /// not the message. Without file access such an attribute is an error
+    /// rather than a garbage datatype.
+    pub fn parse_in_file(
+        data: &[u8],
+        file_data: &[u8],
+        offset_size: u8,
+        length_size: u8,
+    ) -> Result<AttributeMessage, FormatError> {
+        Self::parse_impl(data, length_size, Some((file_data, offset_size)))
+    }
+
+    fn parse_impl(
+        data: &[u8],
+        length_size: u8,
+        file: Option<(&[u8], u8)>,
+    ) -> Result<AttributeMessage, FormatError> {
        ensure_len(data, 0, 2)?;
        let version = data[0];

        match version {
            1 => Self::parse_v1(data, length_size),
-            2 => Self::parse_v2(data, length_size),
-            3 => Self::parse_v3(data, length_size),
+            2 => Self::parse_v2(data, length_size, file),
+            3 => Self::parse_v3(data, length_size, file),
            _ => Err(FormatError::InvalidAttributeVersion(version)),
        }
    }

+    /// The bytes of an embedded datatype/dataspace message, following the
+    /// shared-message reference when `shared` is set.
+    fn embedded_message<'a>(
+        bytes: &'a [u8],
+        shared: bool,
+        msg_type: MessageType,
+        length_size: u8,
+        file: Option<(&[u8], 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(
+            file_data,
+            &shared_ref,
+            msg_type,
+            offset_size,
+            length_size,
+        )
+        .map(Cow::Owned)
+    }
+
    fn parse_v1(data: &[u8], length_size: u8) -> Result<AttributeMessage, FormatError> {
        // version(1) + reserved(1) + name_size(2) + datatype_size(2) + dataspace_size(2) = 8
        ensure_len(data, 0, 8)?;
@@ -94,7 +143,13 @@ impl AttributeMessage {
        })
    }

-    fn parse_v2(data: &[u8], length_size: u8) -> Result<AttributeMessage, FormatError> {
+    fn parse_v2(
+        data: &[u8],
+        length_size: u8,
+        file: Option<(&[u8], u8)>,
+    ) -> Result<AttributeMessage, FormatError> {
+        // Flags: bit 0 = datatype is shared, bit 1 = dataspace is shared.
+        let flags = data.get(1).copied().unwrap_or(0);
        // version(1) + flags(1) + name_size(2) + datatype_size(2) + dataspace_size(2) = 8
        ensure_len(data, 0, 8)?;
        let name_size = u16::from_le_bytes([data[2], data[3]]) as usize;
@@ -110,12 +165,26 @@ impl AttributeMessage {

        // Datatype (NO padding)
        ensure_len(data, pos, datatype_size)?;
-        let (datatype, _) = Datatype::parse(&data[pos..pos + datatype_size])?;
+        let dt_bytes = Self::embedded_message(
+            &data[pos..pos + datatype_size],
+            flags & 0x01 != 0,
+            MessageType::Datatype,
+            length_size,
+            file,
+        )?;
+        let (datatype, _) = Datatype::parse(&dt_bytes)?;
        pos += datatype_size;

        // Dataspace (NO padding)
        ensure_len(data, pos, dataspace_size)?;
-        let dataspace = Dataspace::parse(&data[pos..pos + dataspace_size], length_size)?;
+        let ds_bytes = Self::embedded_message(
+            &data[pos..pos + dataspace_size],
+            flags & 0x02 != 0,
+            MessageType::Dataspace,
+            length_size,
+            file,
+        )?;
+        let dataspace = Dataspace::parse(&ds_bytes, length_size)?;
        pos += dataspace_size;

        let raw_data = compute_raw_data(data, pos, &dataspace, &datatype);
@@ -128,7 +197,13 @@ impl AttributeMessage {
        })
    }

-    fn parse_v3(data: &[u8], length_size: u8) -> Result<AttributeMessage, FormatError> {
+    fn parse_v3(
+        data: &[u8],
+        length_size: u8,
+        file: Option<(&[u8], u8)>,
+    ) -> Result<AttributeMessage, FormatError> {
+        // Flags: bit 0 = datatype is shared, bit 1 = dataspace is shared.
+        let flags = data.get(1).copied().unwrap_or(0);
        // version(1) + flags(1) + name_size(2) + datatype_size(2) + dataspace_size(2) + encoding(1) = 9
        ensure_len(data, 0, 9)?;
        let name_size = u16::from_le_bytes([data[2], data[3]]) as usize;
@@ -145,12 +220,26 @@ impl AttributeMessage {

        // Datatype (NO padding)
        ensure_len(data, pos, datatype_size)?;
-        let (datatype, _) = Datatype::parse(&data[pos..pos + datatype_size])?;
+        let dt_bytes = Self::embedded_message(
+            &data[pos..pos + datatype_size],
+            flags & 0x01 != 0,
+            MessageType::Datatype,
+            length_size,
+            file,
+        )?;
+        let (datatype, _) = Datatype::parse(&dt_bytes)?;
        pos += datatype_size;

        // Dataspace (NO padding)
        ensure_len(data, pos, dataspace_size)?;
-        let dataspace = Dataspace::parse(&data[pos..pos + dataspace_size], length_size)?;
+        let ds_bytes = Self::embedded_message(
+            &data[pos..pos + dataspace_size],
+            flags & 0x02 != 0,
+            MessageType::Dataspace,
+            length_size,
+            file,
+        )?;
+        let dataspace = Dataspace::parse(&ds_bytes, length_size)?;
        pos += dataspace_size;

        let raw_data = compute_raw_data(data, pos, &dataspace, &datatype);
@@ -326,10 +415,20 @@ pub fn extract_attributes_full(
                    offset_size,
                    length_size,
                )?;
-                let attr = AttributeMessage::parse(&resolved_data, length_size)?;
+                let attr = AttributeMessage::parse_in_file(
+                    &resolved_data,
+                    file_data,
+                    offset_size,
+                    length_size,
+                )?;
                attrs.push(attr);
            } else {
-                let attr = AttributeMessage::parse(&msg.data, length_size)?;
+                let attr = AttributeMessage::parse_in_file(
+                    &msg.data,
+                    file_data,
+                    offset_size,
+                    length_size,
+                )?;
                attrs.push(attr);
            }
        }
@@ -399,7 +498,8 @@ fn extract_dense_attributes(
        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(&attr_data, length_size)?;
+        let attr =
+            AttributeMessage::parse_in_file(&attr_data, file_data, offset_size, length_size)?;
        attrs.push(attr);
    }

@@ -472,14 +572,13 @@ mod tests {

        // Name padded to 8 bytes
        data.extend_from_slice(name);
-        while data.len() % 8 != 0 || data.len() == 8 {
+        if data.len() % 8 != 0 || data.len() == 8 {
            // Pad name to 8-byte boundary from start of name
            let name_start = 8;
            let name_padded = pad8(name_size);
            while data.len() < name_start + name_padded {
                data.push(0);
            }
-            break;
        }

        // Datatype padded to 8 bytes
@@ -749,11 +848,11 @@ mod tests {
        data.extend_from_slice(name);
        data.extend_from_slice(&dt_bytes);
        data.extend_from_slice(&ds_bytes);
-        data.extend_from_slice(&3.14f64.to_le_bytes());
+        data.extend_from_slice(&3.25f64.to_le_bytes());

        let attr = AttributeMessage::parse(&data, 8).unwrap();
        let vals = attr.read_as_f64().unwrap();
-        assert_eq!(vals, vec![3.14]);
+        assert_eq!(vals, vec![3.25]);
    }

    #[test]
@@ -416,6 +416,7 @@ fn header_max_total_records(max_leaf_nrec: u64, depth: u16) -> u64 {
 mod tests {
    use super::*;

+    #[allow(clippy::too_many_arguments)]
    fn build_btree_v2_header(
        tree_type: u8,
        node_size: u32,
@@ -374,6 +374,11 @@ impl ChunkCache {

    // ----- Index operations -----

+    /// The most decompressed bytes this cache will hold.
+    pub fn max_bytes(&self) -> usize {
+        self.inner.lock().map(|g| g.max_bytes).unwrap_or(0)
+    }
+
    /// Bind the cache to the dataset at chunk-index address `addr`.
    ///
    /// The cache is shared per file across all of its datasets. If the cache
@@ -132,6 +132,64 @@ fn ensure_len(data: &[u8], offset: usize, needed: usize) -> Result<(), FormatErr
    Ok(())
 }

+/// `elements * elem_size` for sizes that come from the file. Dataspace and
+/// chunk dimensions are untrusted 64-bit fields, so a crafted file can make
+/// the plain product wrap to a small number (or to something enormous).
+pub(crate) fn checked_byte_len(elements: u64, elem_size: usize) -> Result<usize, FormatError> {
+    usize::try_from(elements)
+        .ok()
+        .and_then(|n| n.checked_mul(elem_size))
+        .ok_or_else(|| {
+            FormatError::Overflow(format!(
+                "{elements} elements of {elem_size} bytes exceeds the addressable size"
+            ))
+        })
+}
+
+/// Product of chunk dimensions times the element size, overflow-checked.
+pub(crate) fn checked_chunk_byte_len(
+    chunk_dims: &[usize],
+    elem_size: usize,
+) -> Result<usize, FormatError> {
+    chunk_dims
+        .iter()
+        .try_fold(elem_size, |acc, &d| acc.checked_mul(d))
+        .ok_or_else(|| {
+            FormatError::Overflow(format!(
+                "chunk dimensions {chunk_dims:?} x {elem_size} bytes exceeds the addressable size"
+            ))
+        })
+}
+
+/// A zero-filled output buffer of `len` bytes. `vec![0; len]` aborts the
+/// process when the allocation fails; a size taken from the file must surface
+/// as an error instead.
+pub(crate) fn alloc_output(len: usize) -> Result<Vec<u8>, FormatError> {
+    if len == 0 {
+        return Ok(Vec::new());
+    }
+    let failed =
+        || FormatError::Overflow(format!("cannot allocate {len} bytes for dataset output"));
+    let layout = core::alloc::Layout::array::<u8>(len).map_err(|_| failed())?;
+    // Ask the allocator for zeroed memory instead of reserving and then
+    // writing zeros: for a large buffer the OS hands out already-zero pages
+    // lazily, where an explicit fill touches every page up front — and most of
+    // the buffer is about to be overwritten with chunk data anyway.
+    //
+    // SAFETY (both arms): `layout` has non-zero size (len > 0) and alignment 1.
+    #[cfg(feature = "std")]
+    let ptr = unsafe { std::alloc::alloc_zeroed(layout) };
+    #[cfg(not(feature = "std"))]
+    let ptr = unsafe { alloc::alloc::alloc_zeroed(layout) };
+    if ptr.is_null() {
+        return Err(failed());
+    }
+    // SAFETY: `ptr` came from the global allocator with the layout of
+    // `[u8; len]`, which is exactly what `Vec<u8>` with capacity `len` frees;
+    // all `len` bytes are initialised (zero).
+    Ok(unsafe { Vec::from_raw_parts(ptr, len, len) })
+}
+
 fn read_offset(data: &[u8], pos: usize, size: u8) -> Result<u64, FormatError> {
    let s = size as usize;
    if pos.checked_add(s).is_none_or(|end| end > data.len()) {
@@ -321,15 +379,127 @@ pub fn generate_implicit_chunks(
 }

 /// Read a chunked dataset, decompressing chunks as needed.
-pub fn read_chunked_data(
+/// Chunks decompressed together before being copied out, bounding the extra
+/// memory a parallel full read holds at once.
+const DECODE_BATCH: usize = 128;
+
+/// B-tree v2 record types used for chunk indexing.
+const BT2_CHUNK_UNFILTERED: u8 = 10;
+const BT2_CHUNK_FILTERED: u8 = 11;
+
+/// Chunks indexed by a version-2 B-tree (layout v4, index type 5).
+///
+/// Record layouts (all little endian):
+/// * type 10, unfiltered: address, then one 8-byte *scaled* offset per
+///   dimension (offset / chunk dimension);
+/// * type 11, filtered: address, stored chunk size (a variable number of
+///   bytes), 4-byte filter mask, then the scaled offsets.
+///
+/// The width of the stored-size field depends on the largest possible chunk;
+/// rather than re-derive the library's formula it is taken from the record
+/// size the tree header declares, which is what actually governs the bytes.
+fn read_btree_v2_chunks(
+    file_data: &[u8],
+    addr: u64,
+    chunk_dims: &[usize],
+    elem_size: usize,
+    offset_size: u8,
+    length_size: u8,
+) -> Result<Vec<ChunkInfo>, FormatError> {
+    use crate::btree_v2::{BTreeV2Header, collect_btree_v2_records};
+
+    let bad = |what: &str| FormatError::ChunkedReadError(format!("B-tree v2 chunk index: {what}"));
+    let header = BTreeV2Header::parse(file_data, addr as usize, offset_size, length_size)?;
+    let rank = chunk_dims.len();
+    let os = offset_size as usize;
+    let record_size = header.record_size as usize;
+    let size_len = match header.tree_type {
+        BT2_CHUNK_UNFILTERED => {
+            if record_size != os + 8 * rank {
+                return Err(bad("unexpected record size for unfiltered chunks"));
+            }
+            0
+        }
+        BT2_CHUNK_FILTERED => {
+            let fixed = os + 4 + 8 * rank;
+            let size_len = record_size
+                .checked_sub(fixed)
+                .ok_or_else(|| bad("record too small"))?;
+            if !(1..=8).contains(&size_len) {
+                return Err(bad("implausible chunk-size field width"));
+            }
+            size_len
+        }
+        _ => return Err(bad("tree is not a chunk index")),
+    };
+    let unfiltered_bytes = checked_chunk_byte_len(chunk_dims, elem_size)?;
+    let unfiltered_bytes =
+        u32::try_from(unfiltered_bytes).map_err(|_| bad("chunk larger than 4 GiB"))?;
+
+    let records = collect_btree_v2_records(file_data, &header, offset_size, length_size)?;
+    let mut chunks = Vec::with_capacity(records.len());
+    for record in &records {
+        let data = record.data.as_slice();
+        if data.len() < record_size {
+            return Err(bad("truncated record"));
+        }
+        let address = read_offset(data, 0, offset_size)?;
+        let mut pos = os;
+        let (chunk_size, filter_mask) = if size_len == 0 {
+            (unfiltered_bytes, 0)
+        } else {
+            let mut size = 0u64;
+            for (i, &b) in data[pos..pos + size_len].iter().enumerate() {
+                size |= u64::from(b) << (8 * i);
+            }
+            pos += size_len;
+            let mask = u32::from_le_bytes([data[pos], data[pos + 1], data[pos + 2], data[pos + 3]]);
+            pos += 4;
+            (
+                u32::try_from(size).map_err(|_| bad("stored chunk larger than 4 GiB"))?,
+                mask,
+            )
+        };
+        let mut offsets = Vec::with_capacity(rank);
+        for &dim in chunk_dims {
+            let scaled = u64::from_le_bytes([
+                data[pos],
+                data[pos + 1],
+                data[pos + 2],
+                data[pos + 3],
+                data[pos + 4],
+                data[pos + 5],
+                data[pos + 6],
+                data[pos + 7],
+            ]);
+            pos += 8;
+            offsets.push(
+                scaled
+                    .checked_mul(dim as u64)
+                    .ok_or_else(|| bad("chunk offset overflows"))?,
+            );
+        }
+        chunks.push(ChunkInfo {
+            chunk_size,
+            filter_mask,
+            offsets,
+            address,
+        });
+    }
+    Ok(chunks)
+}
+
+/// Every allocated chunk of a chunked dataset, for any supported chunk index,
+/// plus the spatial chunk dimensions. Chunks the file never allocated (sparse
+/// datasets) are simply absent from the list.
+pub fn list_chunks(
    file_data: &[u8],
    layout: &DataLayout,
    dataspace: &Dataspace,
-    datatype: &Datatype,
-    pipeline: Option<&FilterPipeline>,
+    elem_size: usize,
    offset_size: u8,
    length_size: u8,
-) -> Result<Vec<u8>, FormatError> {
+) -> Result<(Vec<ChunkInfo>, Vec<usize>), FormatError> {
    let (
        chunk_dimensions,
        version,
@@ -363,8 +533,6 @@ pub fn read_chunked_data(
    let addr = addr_opt
        .ok_or_else(|| FormatError::ChunkedReadError("no address for chunked layout".into()))?;

-    let elem_size = datatype.type_size() as usize;
-
    // Both v3 and v4 include element size as last dim (rank+1)
    let ndims = chunk_dimensions.len();
    let rank = ndims
@@ -393,7 +561,7 @@ pub fn read_chunked_data(
        }
        (4, Some(1)) => {
            // Single chunk — one chunk covering the entire dataset
-            let chunk_byte_size: usize = chunk_dims.iter().product::<usize>() * elem_size;
+            let chunk_byte_size = checked_chunk_byte_len(&chunk_dims, elem_size)?;
            let (csize, fmask) = if let Some(fs) = single_filtered_size {
                (fs as u32, single_filter_mask.unwrap_or(0))
            } else {
@@ -446,6 +614,18 @@ pub fn read_chunked_data(
                length_size,
            )?
        }
+        (4, Some(5)) => {
+            // Version-2 B-tree: what the library uses for a dataset with two
+            // or more unlimited dimensions.
+            read_btree_v2_chunks(
+                file_data,
+                addr,
+                &chunk_dims,
+                elem_size,
+                offset_size,
+                length_size,
+            )?
+        }
        (v, idx) => {
            return Err(FormatError::ChunkedReadError(format!(
                "unsupported chunked layout version={v}, index_type={idx:?}"
@@ -453,10 +633,38 @@ pub fn read_chunked_data(
        }
    };

+    Ok((chunks, chunk_dims))
+}
+
+pub fn read_chunked_data(
+    file_data: &[u8],
+    layout: &DataLayout,
+    dataspace: &Dataspace,
+    datatype: &Datatype,
+    pipeline: Option<&FilterPipeline>,
+    offset_size: u8,
+    length_size: u8,
+) -> Result<Vec<u8>, FormatError> {
+    let elem_size = datatype.type_size() as usize;
+    let (chunks, chunk_dims) = list_chunks(
+        file_data,
+        layout,
+        dataspace,
+        elem_size,
+        offset_size,
+        length_size,
+    )?;
+    let rank = chunk_dims.len();
+    let ds_dims: Vec<usize> = dataspace.dimensions.iter().map(|&d| d as usize).collect();
+
    // Assemble output
-    let total_elements = dataspace.num_elements() as usize;
-    let total_bytes = total_elements * elem_size;
-    let mut output = vec![0u8; total_bytes];
+    let total_bytes = checked_byte_len(dataspace.checked_num_elements()?, elem_size)?;
+    if total_bytes == 0 {
+        // Also keeps the stride products below in range: with a zero-sized
+        // dimension the total is 0 even if other dimensions are huge.
+        return Ok(Vec::new());
+    }
+    let mut output = alloc_output(total_bytes)?;

    let mut ds_strides = vec![1usize; rank];
    for i in (0..rank.saturating_sub(1)).rev() {
@@ -468,8 +676,7 @@ pub fn read_chunked_data(
        chunk_strides[i] = chunk_strides[i + 1] * chunk_dims[i + 1];
    }

-    let chunk_total_elements: usize = chunk_dims.iter().product();
-    let chunk_total_bytes = chunk_total_elements * elem_size;
+    let chunk_total_bytes = checked_chunk_byte_len(&chunk_dims, elem_size)?;

    // Fast path: no filters — copy directly from file_data without intermediate alloc
    if pipeline.is_none() {
@@ -561,29 +768,12 @@ pub fn read_chunked_data_cached(
    length_size: u8,
    cache: &ChunkCache,
 ) -> Result<Vec<u8>, FormatError> {
-    let (
-        chunk_dimensions,
-        version,
-        chunk_index_type,
-        addr_opt,
-        single_filtered_size,
-        single_filter_mask,
-    ) = match layout {
+    let (chunk_dimensions, addr_opt) = match layout {
        DataLayout::Chunked {
            chunk_dimensions,
            btree_address,
-            version,
-            chunk_index_type,
-            single_chunk_filtered_size,
-            single_chunk_filter_mask,
-        } => (
-            chunk_dimensions,
-            *version,
-            *chunk_index_type,
-            *btree_address,
-            *single_chunk_filtered_size,
-            *single_chunk_filter_mask,
-        ),
+            ..
+        } => (chunk_dimensions, *btree_address),
        _ => {
            return Err(FormatError::ChunkedReadError(
                "expected chunked layout".into(),
@@ -620,78 +810,27 @@ pub fn read_chunked_data_cached(

    // Populate chunk index on first access
    if !cache.has_index() {
-        let chunks = match (version, chunk_index_type) {
-            (3, _) => collect_chunk_info(file_data, addr, ndims, offset_size, length_size)?,
-            (4, Some(1)) => {
-                let chunk_byte_size: usize = chunk_dims.iter().product::<usize>() * elem_size;
-                let (csize, fmask) = if let Some(fs) = single_filtered_size {
-                    (fs as u32, single_filter_mask.unwrap_or(0))
-                } else {
-                    (chunk_byte_size as u32, 0)
-                };
-                vec![ChunkInfo {
-                    chunk_size: csize,
-                    filter_mask: fmask,
-                    offsets: vec![0u64; rank],
-                    address: addr,
-                }]
-            }
-            (4, Some(2)) => {
-                let spatial_chunk_dims: &[u32] = &chunk_dimensions[..rank];
-                generate_implicit_chunks(
-                    addr,
-                    &dataspace.dimensions,
-                    spatial_chunk_dims,
-                    elem_size as u32,
-                )
-            }
-            (4, Some(3)) => {
-                let spatial_chunk_dims: &[u32] = &chunk_dimensions[..rank];
-                let header =
-                    FixedArrayHeader::parse(file_data, addr as usize, offset_size, length_size)?;
-                read_fixed_array_chunks(
-                    file_data,
-                    &header,
-                    &dataspace.dimensions,
-                    spatial_chunk_dims,
-                    elem_size as u32,
-                    offset_size,
-                    length_size,
-                )?
-            }
-            (4, Some(4)) => {
-                let spatial_chunk_dims: &[u32] = &chunk_dimensions[..rank];
-                let header = ExtensibleArrayHeader::parse(
-                    file_data,
-                    addr as usize,
-                    offset_size,
-                    length_size,
-                )?;
-                read_extensible_array_chunks(
-                    file_data,
-                    &header,
-                    &dataspace.dimensions,
-                    spatial_chunk_dims,
-                    elem_size as u32,
-                    offset_size,
-                    length_size,
-                )?
-            }
-            (v, idx) => {
-                return Err(FormatError::ChunkedReadError(format!(
-                    "unsupported chunked layout version={v}, index_type={idx:?}"
-                )));
-            }
-        };
+        let (chunks, _) = list_chunks(
+            file_data,
+            layout,
+            dataspace,
+            elem_size,
+            offset_size,
+            length_size,
+        )?;
        cache.populate_index(&chunks, rank);
    }

    let chunks = cache.all_indexed_chunks().unwrap_or_default();

    // Assemble output
-    let total_elements = dataspace.num_elements() as usize;
-    let total_bytes = total_elements * elem_size;
-    let mut output = vec![0u8; total_bytes];
+    let total_bytes = checked_byte_len(dataspace.checked_num_elements()?, elem_size)?;
+    if total_bytes == 0 {
+        // Also keeps the stride products below in range: with a zero-sized
+        // dimension the total is 0 even if other dimensions are huge.
+        return Ok(Vec::new());
+    }
+    let mut output = alloc_output(total_bytes)?;

    let mut ds_strides = vec![1usize; rank];
    for i in (0..rank.saturating_sub(1)).rev() {
@@ -703,55 +842,88 @@ pub fn read_chunked_data_cached(
        chunk_strides[i] = chunk_strides[i + 1] * chunk_dims[i + 1];
    }

-    let chunk_total_elements: usize = chunk_dims.iter().product();
-    let chunk_total_bytes = chunk_total_elements * elem_size;
-
-    for chunk_info in &chunks {
-        let coord: Vec<u64> = chunk_info.offsets.iter().take(rank).copied().collect();
-
-        // Try decompressed cache first
-        let decompressed = if let Some(cached) = cache.get_decompressed_aligned(&coord) {
-            cached
-        } else {
-            // Decompress from file
-            let c_addr = chunk_info.address as usize;
-            let size = chunk_info.chunk_size as usize;
-            ensure_len(file_data, c_addr, size)?;
-            let raw_chunk = &file_data[c_addr..c_addr + size];
-            let dec = if let Some(pl) = pipeline {
-                if chunk_info.filter_mask == 0 {
-                    decompress_chunk(raw_chunk, pl, chunk_total_bytes, elem_size as u32)?
-                } else {
-                    raw_chunk.to_vec()
-                }
-            } else {
-                raw_chunk.to_vec()
-            };
-            cache.put_decompressed(coord, dec)
-        };
+    let chunk_total_bytes = checked_chunk_byte_len(&chunk_dims, elem_size)?;

+    let mut place = |data: &[u8], chunk_info: &ChunkInfo| {
+        if rank == 0 {
+            let copy_len = data.len().min(output.len());
+            output[..copy_len].copy_from_slice(&data[..copy_len]);
+            return;
+        }
        let chunk_offsets: Vec<usize> = chunk_info
            .offsets
            .iter()
            .take(rank)
            .map(|&o| o as usize)
            .collect();
+        copy_chunk_to_output(
+            data,
+            &mut output,
+            &chunk_offsets,
+            &chunk_dims,
+            &ds_dims,
+            &ds_strides,
+            &chunk_strides,
+            elem_size,
+            rank,
+        );
+    };
+    let raw_bytes = |chunk_info: &ChunkInfo| -> Result<&[u8], FormatError> {
+        let c_addr = chunk_info.address as usize;
+        let size = chunk_info.chunk_size as usize;
+        ensure_len(file_data, c_addr, size)?;
+        Ok(&file_data[c_addr..c_addr + size])
+    };

-        if rank == 0 {
-            let copy_len = decompressed.len().min(output.len());
-            output[..copy_len].copy_from_slice(&decompressed[..copy_len]);
-        } else {
-            copy_chunk_to_output(
-                &decompressed,
-                &mut output,
-                &chunk_offsets,
-                &chunk_dims,
-                &ds_dims,
-                &ds_strides,
-                &chunk_strides,
-                elem_size,
-                rank,
-            );
+    // Chunks stored as-is (no pipeline, or the filter mask says this chunk
+    // skipped it) are copied straight from the file bytes: they are already in
+    // memory, so routing them through a Vec and then an aligned cache buffer
+    // was two extra copies of the whole dataset for nothing.
+    let stored_raw = |c: &ChunkInfo| pipeline.is_none() || c.filter_mask != 0;
+    let mut misses: Vec<&ChunkInfo> = Vec::new();
+    for chunk_info in &chunks {
+        if stored_raw(chunk_info) {
+            place(raw_bytes(chunk_info)?, chunk_info);
+            continue;
+        }
+        let coord: Vec<u64> = chunk_info.offsets.iter().take(rank).copied().collect();
+        match cache.get_decompressed_aligned(&coord) {
+            Some(cached) => place(&cached, chunk_info),
+            None => misses.push(chunk_info),
+        }
+    }
+
+    // Decompress what the cache didn't have, a bounded batch at a time — in
+    // parallel with the `parallel` feature (this path, the one the facade
+    // uses, was sequential; only the uncached reader was parallel). Chunks are
+    // cached only when the whole dataset fits: pushing a larger dataset
+    // through the cache just evicts each chunk moments after inserting it.
+    let cache_them = total_bytes <= cache.max_bytes();
+    if let Some(pl) = pipeline {
+        let decode = |c: &&ChunkInfo| -> Result<Vec<u8>, FormatError> {
+            decompress_chunk(raw_bytes(c)?, pl, chunk_total_bytes, elem_size as u32)
+        };
+        for batch in misses.chunks(DECODE_BATCH) {
+            #[cfg(feature = "parallel")]
+            let decoded: Vec<Result<Vec<u8>, FormatError>> = if batch.len() >= 4 {
+                use rayon::prelude::*;
+                batch.par_iter().map(decode).collect()
+            } else {
+                batch.iter().map(decode).collect()
+            };
+            #[cfg(not(feature = "parallel"))]
+            let decoded: Vec<Result<Vec<u8>, FormatError>> = batch.iter().map(decode).collect();
+
+            for (chunk_info, data) in batch.iter().zip(decoded) {
+                let data = data?;
+                if cache_them {
+                    let coord: Vec<u64> = chunk_info.offsets.iter().take(rank).copied().collect();
+                    let cached = cache.put_decompressed(coord, data);
+                    place(&cached, chunk_info);
+                } else {
+                    place(&data, chunk_info);
+                }
+            }
        }
    }

@@ -914,29 +1086,12 @@ pub fn read_chunked_data_sweep(
    cache: &ChunkCache,
    sweep: &mut SweepContext,
 ) -> Result<Vec<u8>, FormatError> {
-    let (
-        chunk_dimensions,
-        version,
-        chunk_index_type,
-        addr_opt,
-        single_filtered_size,
-        single_filter_mask,
-    ) = match layout {
+    let (chunk_dimensions, addr_opt) = match layout {
        DataLayout::Chunked {
            chunk_dimensions,
            btree_address,
-            version,
-            chunk_index_type,
-            single_chunk_filtered_size,
-            single_chunk_filter_mask,
-        } => (
-            chunk_dimensions,
-            *version,
-            *chunk_index_type,
-            *btree_address,
-            *single_chunk_filtered_size,
-            *single_chunk_filter_mask,
-        ),
+            ..
+        } => (chunk_dimensions, *btree_address),
        _ => {
            return Err(FormatError::ChunkedReadError(
                "expected chunked layout".into(),
@@ -973,78 +1128,27 @@ pub fn read_chunked_data_sweep(

    // Populate chunk index on first access
    if !cache.has_index() {
-        let chunks = match (version, chunk_index_type) {
-            (3, _) => collect_chunk_info(file_data, addr, ndims, offset_size, length_size)?,
-            (4, Some(1)) => {
-                let chunk_byte_size: usize = chunk_dims.iter().product::<usize>() * elem_size;
-                let (csize, fmask) = if let Some(fs) = single_filtered_size {
-                    (fs as u32, single_filter_mask.unwrap_or(0))
-                } else {
-                    (chunk_byte_size as u32, 0)
-                };
-                vec![ChunkInfo {
-                    chunk_size: csize,
-                    filter_mask: fmask,
-                    offsets: vec![0u64; rank],
-                    address: addr,
-                }]
-            }
-            (4, Some(2)) => {
-                let spatial_chunk_dims: &[u32] = &chunk_dimensions[..rank];
-                generate_implicit_chunks(
-                    addr,
-                    &dataspace.dimensions,
-                    spatial_chunk_dims,
-                    elem_size as u32,
-                )
-            }
-            (4, Some(3)) => {
-                let spatial_chunk_dims: &[u32] = &chunk_dimensions[..rank];
-                let header =
-                    FixedArrayHeader::parse(file_data, addr as usize, offset_size, length_size)?;
-                read_fixed_array_chunks(
-                    file_data,
-                    &header,
-                    &dataspace.dimensions,
-                    spatial_chunk_dims,
-                    elem_size as u32,
-                    offset_size,
-                    length_size,
-                )?
-            }
-            (4, Some(4)) => {
-                let spatial_chunk_dims: &[u32] = &chunk_dimensions[..rank];
-                let header = ExtensibleArrayHeader::parse(
-                    file_data,
-                    addr as usize,
-                    offset_size,
-                    length_size,
-                )?;
-                read_extensible_array_chunks(
-                    file_data,
-                    &header,
-                    &dataspace.dimensions,
-                    spatial_chunk_dims,
-                    elem_size as u32,
-                    offset_size,
-                    length_size,
-                )?
-            }
-            (v, idx) => {
-                return Err(FormatError::ChunkedReadError(format!(
-                    "unsupported chunked layout version={v}, index_type={idx:?}"
-                )));
-            }
-        };
+        let (chunks, _) = list_chunks(
+            file_data,
+            layout,
+            dataspace,
+            elem_size,
+            offset_size,
+            length_size,
+        )?;
        cache.populate_index(&chunks, rank);
    }

    let chunks = cache.all_indexed_chunks().unwrap_or_default();

    // Assemble output
-    let total_elements = dataspace.num_elements() as usize;
-    let total_bytes = total_elements * elem_size;
-    let mut output = vec![0u8; total_bytes];
+    let total_bytes = checked_byte_len(dataspace.checked_num_elements()?, elem_size)?;
+    if total_bytes == 0 {
+        // Also keeps the stride products below in range: with a zero-sized
+        // dimension the total is 0 even if other dimensions are huge.
+        return Ok(Vec::new());
+    }
+    let mut output = alloc_output(total_bytes)?;

    let mut ds_strides = vec![1usize; rank];
    for i in (0..rank.saturating_sub(1)).rev() {
@@ -1056,8 +1160,7 @@ pub fn read_chunked_data_sweep(
        chunk_strides[i] = chunk_strides[i + 1] * chunk_dims[i + 1];
    }

-    let chunk_total_elements: usize = chunk_dims.iter().product();
-    let chunk_total_bytes = chunk_total_elements * elem_size;
+    let chunk_total_bytes = checked_chunk_byte_len(&chunk_dims, elem_size)?;

    for chunk_info in &chunks {
        let coord: Vec<u64> = chunk_info.offsets.iter().take(rank).copied().collect();
@@ -1137,29 +1240,12 @@ pub fn read_chunked_data_indexed(
    length_size: u8,
    cache: &ChunkCache,
 ) -> Result<Vec<u8>, FormatError> {
-    let (
-        chunk_dimensions,
-        version,
-        chunk_index_type,
-        addr_opt,
-        single_filtered_size,
-        single_filter_mask,
-    ) = match layout {
+    let (chunk_dimensions, addr_opt) = match layout {
        DataLayout::Chunked {
            chunk_dimensions,
            btree_address,
-            version,
-            chunk_index_type,
-            single_chunk_filtered_size,
-            single_chunk_filter_mask,
-        } => (
-            chunk_dimensions,
-            *version,
-            *chunk_index_type,
-            *btree_address,
-            *single_chunk_filtered_size,
-            *single_chunk_filter_mask,
-        ),
+            ..
+        } => (chunk_dimensions, *btree_address),
        _ => {
            return Err(FormatError::ChunkedReadError(
                "expected chunked layout".into(),
@@ -1196,69 +1282,14 @@ pub fn read_chunked_data_indexed(

    // Build chunk index on first access
    if !cache.has_chunk_index() {
-        let chunks = match (version, chunk_index_type) {
-            (3, _) => collect_chunk_info(file_data, addr, ndims, offset_size, length_size)?,
-            (4, Some(1)) => {
-                let chunk_byte_size: usize = chunk_dims.iter().product::<usize>() * elem_size;
-                let (csize, fmask) = if let Some(fs) = single_filtered_size {
-                    (fs as u32, single_filter_mask.unwrap_or(0))
-                } else {
-                    (chunk_byte_size as u32, 0)
-                };
-                vec![ChunkInfo {
-                    chunk_size: csize,
-                    filter_mask: fmask,
-                    offsets: vec![0u64; rank],
-                    address: addr,
-                }]
-            }
-            (4, Some(2)) => {
-                let spatial_chunk_dims: &[u32] = &chunk_dimensions[..rank];
-                generate_implicit_chunks(
-                    addr,
-                    &dataspace.dimensions,
-                    spatial_chunk_dims,
-                    elem_size as u32,
-                )
-            }
-            (4, Some(3)) => {
-                let spatial_chunk_dims: &[u32] = &chunk_dimensions[..rank];
-                let header =
-                    FixedArrayHeader::parse(file_data, addr as usize, offset_size, length_size)?;
-                read_fixed_array_chunks(
-                    file_data,
-                    &header,
-                    &dataspace.dimensions,
-                    spatial_chunk_dims,
-                    elem_size as u32,
-                    offset_size,
-                    length_size,
-                )?
-            }
-            (4, Some(4)) => {
-                let spatial_chunk_dims: &[u32] = &chunk_dimensions[..rank];
-                let header = ExtensibleArrayHeader::parse(
-                    file_data,
-                    addr as usize,
-                    offset_size,
-                    length_size,
-                )?;
-                read_extensible_array_chunks(
-                    file_data,
-                    &header,
-                    &dataspace.dimensions,
-                    spatial_chunk_dims,
-                    elem_size as u32,
-                    offset_size,
-                    length_size,
-                )?
-            }
-            (v, idx) => {
-                return Err(FormatError::ChunkedReadError(format!(
-                    "unsupported chunked layout version={v}, index_type={idx:?}"
-                )));
-            }
-        };
+        let (chunks, _) = list_chunks(
+            file_data,
+            layout,
+            dataspace,
+            elem_size,
+            offset_size,
+            length_size,
+        )?;
        cache.populate_chunk_index(&chunks, rank);
        // Also populate the legacy index for compatibility
        if !cache.has_index() {
@@ -1463,6 +1494,64 @@ fn copy_chunk_to_output(
 mod tests {
    use super::*;

+    fn simple_space(dimensions: Vec<u64>) -> Dataspace {
+        Dataspace {
+            space_type: crate::dataspace::DataspaceType::Simple,
+            rank: dimensions.len() as u8,
+            dimensions,
+            max_dimensions: None,
+        }
+    }
+
+    #[test]
+    fn crafted_dimensions_are_errors_not_wraparound() {
+        // 2^63 * 2 wraps to 0 with a plain product; 2^40 * 2^40 wraps too.
+        for dims in [
+            vec![1u64 << 63, 2],
+            vec![1 << 40, 1 << 40],
+            vec![u64::MAX, u64::MAX],
+        ] {
+            let space = simple_space(dims.clone());
+            assert!(
+                matches!(space.checked_num_elements(), Err(FormatError::Overflow(_))),
+                "{dims:?}"
+            );
+            // The infallible accessor saturates instead of wrapping.
+            assert_eq!(space.num_elements(), u64::MAX, "{dims:?}");
+        }
+        assert_eq!(simple_space(vec![3, 4]).checked_num_elements().unwrap(), 12);
+        // A zero-sized dimension makes the whole product 0, not an overflow.
+        assert_eq!(
+            simple_space(vec![0, 1 << 40, 1 << 40])
+                .checked_num_elements()
+                .unwrap(),
+            0
+        );
+    }
+
+    #[test]
+    fn byte_length_helpers_check_overflow() {
+        assert_eq!(checked_byte_len(10, 8).unwrap(), 80);
+        assert!(matches!(
+            checked_byte_len(u64::MAX, 8),
+            Err(FormatError::Overflow(_))
+        ));
+        assert_eq!(checked_chunk_byte_len(&[10, 10], 4).unwrap(), 400);
+        assert!(matches!(
+            checked_chunk_byte_len(&[usize::MAX, 2], 4),
+            Err(FormatError::Overflow(_))
+        ));
+    }
+
+    #[test]
+    fn unallocatable_output_is_an_error_not_an_abort() {
+        assert_eq!(alloc_output(16).unwrap(), vec![0u8; 16]);
+        assert!(matches!(
+            alloc_output(usize::MAX / 2),
+            Err(FormatError::Overflow(_))
+        ));
+    }
+
    fn write_offset(buf: &mut Vec<u8>, val: u64, size: u8) {
        match size {
            4 => buf.extend_from_slice(&(val as u32).to_le_bytes()),
@@ -1657,9 +1746,9 @@ mod tests {
            let chunk_bytes = chunk_size_elems * elem_size; // full chunk allocation

            // Write chunk data (full chunk size, padding with zeros)
-            for i in start..end {
+            for (i, value) in values.iter().enumerate().take(end).skip(start) {
                let byte_offset = data_offset + (i - start) * elem_size;
-                file_data[byte_offset..byte_offset + 8].copy_from_slice(&values[i].to_le_bytes());
+                file_data[byte_offset..byte_offset + 8].copy_from_slice(&value.to_le_bytes());
            }

            chunk_infos.push(ChunkInfo {
@@ -1837,8 +1926,8 @@ mod tests {
        for chunk_idx in 0..2 {
            let start = chunk_idx * chunk_elems;
            let mut chunk_bytes = Vec::new();
-            for i in start..start + chunk_elems {
-                chunk_bytes.extend_from_slice(&values[i].to_le_bytes());
+            for value in values.iter().skip(start).take(chunk_elems) {
+                chunk_bytes.extend_from_slice(&value.to_le_bytes());
            }
            let compressed = compress_chunk(&chunk_bytes, &pipeline, elem_size as u32).unwrap();

@@ -2152,21 +2241,23 @@ mod tests {
    }

    #[test]
-    fn cached_read_second_call_uses_cache() {
+    fn cached_read_second_call_reuses_the_index() {
        let values: Vec<f64> = (0..20).map(|i| i as f64).collect();
        let (file_data, layout, dataspace) = build_1d_chunked_file(&values, 10);
        let datatype = make_f64_type();
        let cache = ChunkCache::new();

-        // First read — populates index + decompressed cache
+        // First read — populates the chunk index. These chunks are stored
+        // unfiltered, so they are copied straight from the file bytes and the
+        // decompressed-chunk cache is (deliberately) not involved.
        let raw1 = read_chunked_data_cached(
            &file_data, &layout, &dataspace, &datatype, None, 8, 8, &cache,
        )
        .unwrap();
        assert!(cache.has_index());
-        assert!(cache.cached_chunk_count() > 0);
+        assert_eq!(cache.cached_chunk_count(), 0);

-        // Second read — should hit the decompressed cache
+        // Second read — reuses the cached index
        let raw2 = read_chunked_data_cached(
            &file_data, &layout, &dataspace, &datatype, None, 8, 8, &cache,
        )
@@ -15,7 +15,6 @@ use crate::filter_pipeline::{
    FilterDescription, FilterPipeline,
 };
 use crate::filters::compress_chunk;
-
 /// Round a file offset up to the next cache-line boundary.
 ///
 /// This ensures chunk data starts at an address that is a multiple of the
@@ -49,6 +48,38 @@ pub struct ChunkOptions {
    pub pcodec: bool,
 }

+/// Largest chunk the automatic choice produces, in bytes.
+const AUTO_CHUNK_TARGET_BYTES: u64 = 1 << 20;
+
+/// Extent assumed for a dimension that is currently empty (an unlimited
+/// dimension not yet written to) — the same stand-in h5py uses.
+const AUTO_CHUNK_EMPTY_DIM: u64 = 1024;
+
+/// Choose chunk dimensions for a dataset nobody specified them for.
+///
+/// Asking for compression (or any filter) without chunk dimensions used to
+/// make the whole dataset one chunk. That defeats the point of chunking: any
+/// read — even a single row — must decompress everything, and a large dataset
+/// cannot be decompressed in parallel. Datasets up to the target size stay a
+/// single chunk, exactly as before; larger ones are split by halving the
+/// dimensions in turn (so chunks keep roughly the dataset's proportions, the
+/// approach h5py takes) until a chunk fits the target.
+pub fn auto_chunk_dims(shape: &[u64], elem_size: usize) -> Vec<u64> {
+    let mut dims: Vec<u64> = shape
+        .iter()
+        .map(|&d| if d == 0 { AUTO_CHUNK_EMPTY_DIM } else { d })
+        .collect();
+    let elem = elem_size.max(1) as u64;
+    let bytes = |dims: &[u64]| dims.iter().fold(elem, |acc, &d| acc.saturating_mul(d));
+    let mut axis = 0;
+    while bytes(&dims) > AUTO_CHUNK_TARGET_BYTES && dims.iter().any(|&d| d > 1) {
+        let i = axis % dims.len();
+        dims[i] = dims[i].div_ceil(2);
+        axis += 1;
+    }
+    dims
+}
+
 impl ChunkOptions {
    /// Whether any chunking option is enabled.
    pub fn is_chunked(&self) -> bool {
@@ -135,11 +166,17 @@ impl ChunkOptions {

    /// Determine chunk dimensions, using user-specified or auto-computing.
    pub fn resolve_chunk_dims(&self, shape: &[u64]) -> Vec<u64> {
-        if let Some(ref dims) = self.chunk_dims {
-            dims.clone()
-        } else {
-            // Auto chunk: use the full dataset shape (single chunk)
-            shape.to_vec()
+        // Without the element size, assume 8 bytes (the widest common scalar);
+        // the writer uses `resolve_chunk_dims_for`.
+        self.resolve_chunk_dims_for(shape, 8)
+    }
+
+    /// Chunk dimensions for a dataset of `shape` whose elements are `elem_size`
+    /// bytes: the caller's if given, otherwise chosen automatically.
+    pub fn resolve_chunk_dims_for(&self, shape: &[u64], elem_size: usize) -> Vec<u64> {
+        match self.chunk_dims {
+            Some(ref dims) => dims.clone(),
+            None => auto_chunk_dims(shape, elem_size),
        }
    }
 }
@@ -890,6 +927,7 @@ pub fn write_selection_to_buffer(

 #[cfg(test)]
 mod tests {
+
    use super::*;
    use crate::chunked_read::read_chunked_data;
    use crate::data_layout::DataLayout;
@@ -1143,6 +1181,45 @@ mod tests {
        assert_eq!(dims, vec![100, 50]);
    }

+    #[test]
+    fn auto_chunking_splits_only_large_datasets() {
+        let bytes = |dims: &[u64], elem: u64| dims.iter().product::<u64>() * elem;
+        // Up to the target: one chunk, as before.
+        assert_eq!(auto_chunk_dims(&[100, 50], 8), [100, 50]);
+        assert_eq!(auto_chunk_dims(&[131_072], 8), [131_072]); // exactly 1 MiB
+        // Larger: split, keeping proportions, never above the target.
+        let big = auto_chunk_dims(&[4096, 2048], 8);
+        assert!(bytes(&big, 8) <= AUTO_CHUNK_TARGET_BYTES, "{big:?}");
+        assert!(bytes(&big, 8) > AUTO_CHUNK_TARGET_BYTES / 4, "{big:?}");
+        assert_eq!(big[0] / big[1], 2, "proportions kept: {big:?}");
+        // Every dimension stays within the dataset and at least 1.
+        for shape in [
+            vec![10_000_000u64],
+            vec![3, 5_000_000],
+            vec![1, 1, 9_000_000],
+            vec![7; 9],
+        ] {
+            let dims = auto_chunk_dims(&shape, 4);
+            assert!(
+                dims.iter().zip(&shape).all(|(c, s)| *c >= 1 && c <= s),
+                "{shape:?} -> {dims:?}"
+            );
+            assert!(
+                bytes(&dims, 4) <= AUTO_CHUNK_TARGET_BYTES,
+                "{shape:?} -> {dims:?}"
+            );
+        }
+        // An empty (unlimited, unwritten) dimension still gets a usable chunk.
+        let growable = auto_chunk_dims(&[0, 128], 8);
+        assert!(growable[0] >= 1 && bytes(&growable, 8) <= AUTO_CHUNK_TARGET_BYTES);
+        // Explicit dimensions always win.
+        let explicit = ChunkOptions {
+            chunk_dims: Some(vec![10, 10]),
+            ..Default::default()
+        };
+        assert_eq!(explicit.resolve_chunk_dims_for(&[4096, 2048], 8), [10, 10]);
+    }
+
    #[test]
    fn chunk_options_pipeline_deflate() {
        // Auto-shuffle is applied before compression by default (matches h5py).
@@ -1435,9 +1512,20 @@ mod tests {

    // ---- h5py round-trip tests for chunked writes ----

+    /// The Python interpreter to drive interop checks with.
+    ///
+    /// `CLAWHDF5_PYTHON` lets these run against a virtualenv holding h5py,
+    /// which on a PEP 668 "externally managed" system is the only place it
+    /// can be installed. Without it the suite silently skips, and a silent
+    /// skip here is how a datatype bug once reached a release.
+    #[cfg(feature = "std")]
+    fn python() -> String {
+        std::env::var("CLAWHDF5_PYTHON").unwrap_or_else(|_| "python3".to_string())
+    }
+
    #[cfg(feature = "std")]
    fn h5py_available() -> bool {
-        std::process::Command::new("python3")
+        std::process::Command::new(python())
            .args(["-c", "import h5py"])
            .output()
            .map(|o| o.status.success())
@@ -1449,10 +1537,10 @@ mod tests {
        if !h5py_available() {
            panic!("h5py not installed — skipping interop test");
        }
-        let o = std::process::Command::new("python3")
+        let o = std::process::Command::new(python())
            .args(["-c", script])
            .output()
-            .expect("python3");
+            .expect("python interpreter");
        if !o.status.success() {
            panic!("h5py: {}", String::from_utf8_lossy(&o.stderr));
        }
@@ -307,6 +307,24 @@ pub fn read_raw_data_selection(
 ) -> Result<Vec<u8>, FormatError> {
    use crate::selection::Selection;

+    crate::partial_read::validate(selection, &dataspace.dimensions)?;
+
+    // Read only what the selection's bounding box touches when that is
+    // possible; everything below is the decode-everything-then-pick path,
+    // kept for the cases `partial_read` declines.
+    if let Some(selected) = crate::partial_read::read_selection(
+        file_data,
+        layout,
+        dataspace,
+        datatype.type_size() as usize,
+        pipeline,
+        offset_size,
+        length_size,
+        selection,
+    )? {
+        return Ok(selected);
+    }
+
    match selection {
        Selection::All => {
            return read_raw_data_full(
@@ -475,8 +493,10 @@ fn read_virtual_data(
    use crate::selection::Selection;

    let elem_size = datatype.type_size() as usize;
-    let total_elems = dataspace.num_elements() as usize;
-    let mut out = vec![0u8; total_elems.saturating_mul(elem_size)];
+    let mut out = crate::chunked_read::alloc_output(crate::chunked_read::checked_byte_len(
+        dataspace.checked_num_elements()?,
+        elem_size,
+    )?)?;

    let virtual_dims = &dataspace.dimensions;

@@ -598,7 +618,7 @@ fn read_named_dataset_raw(
 }

 /// Extract selected elements from a full dataset buffer.
-fn extract_selection_from_buffer(
+pub fn extract_selection_from_buffer(
    full_data: &[u8],
    dims: &[u64],
    elem_size: usize,
@@ -616,12 +636,14 @@ fn extract_selection_from_buffer(
            block,
        } => {
            let rank = dims.len();
-            let output_elements: usize = count
+            let output_elements = count
                .iter()
                .zip(block.iter())
-                .map(|(&c, &b)| (c * b) as usize)
-                .product();
-            let mut output = vec![0u8; output_elements * elem_size];
+                .try_fold(1u64, |acc, (&c, &b)| acc.checked_mul(c.checked_mul(b)?))
+                .ok_or_else(|| FormatError::Overflow("hyperslab count x block overflows".into()))?;
+            let mut output = crate::chunked_read::alloc_output(
+                crate::chunked_read::checked_byte_len(output_elements, elem_size)?,
+            )?;

            // Compute dataset strides (row-major)
            let mut ds_strides = vec![1usize; rank];
@@ -854,6 +876,30 @@ fn get_size(dt: &Datatype) -> usize {
    dt.type_size() as usize
 }

+/// Reinterpret little-endian bytes as `count` native values of `T` on a
+/// little-endian target, in one copy.
+///
+/// The buffer is allocated uninitialised and filled by the copy. It used to be
+/// `vec![0; count]` first, which for a large dataset meant writing every page
+/// twice (zero it, then overwrite it) — about as expensive as the copy itself.
+#[cfg(target_endian = "little")]
+fn native_le_to_vec<T: Copy>(raw: &[u8], count: usize) -> Vec<T> {
+    let bytes = count * core::mem::size_of::<T>();
+    debug_assert!(bytes <= raw.len());
+    let mut result: Vec<T> = Vec::with_capacity(count);
+    // SAFETY: `result` has capacity for `count` values of `T`, i.e. `bytes`
+    // bytes; `raw` holds at least `bytes` bytes (callers derive `count` from
+    // `raw.len() / size_of::<T>()`); the regions cannot overlap because
+    // `result` was just allocated. Every `T` used here (f32/f64/i32/i64) is
+    // valid for any bit pattern, so after the copy all `count` values are
+    // initialised and `set_len` is sound.
+    unsafe {
+        core::ptr::copy_nonoverlapping(raw.as_ptr(), result.as_mut_ptr().cast::<u8>(), bytes);
+        result.set_len(count);
+    }
+    result
+}
+
 /// Convert raw bytes to `f64` values.
 pub fn read_as_f64(raw: &[u8], datatype: &Datatype) -> Result<Vec<f64>, FormatError> {
    // Array datatypes (e.g. an array-typed compound member) are read as a flat
@@ -881,14 +927,7 @@ pub fn read_as_f64(raw: &[u8], datatype: &Datatype) -> Result<Vec<f64>, FormatEr
            ..
        }
    ) {
-        let mut result = vec![0.0f64; count];
-        // SAFETY: On LE platforms, f64 in-memory representation matches LE bytes.
-        // We copy raw bytes directly into the f64 buffer.
-        // SAFETY: The byte slice is properly aligned for this type and the length is divisible by size_of::<T>().
-        unsafe {
-            core::ptr::copy_nonoverlapping(raw.as_ptr(), result.as_mut_ptr() as *mut u8, raw.len());
-        }
-        return Ok(result);
+        return Ok(native_le_to_vec::<f64>(raw, count));
    }

    let order = get_byte_order(datatype);
@@ -971,12 +1010,7 @@ pub fn read_as_i64(raw: &[u8], datatype: &Datatype) -> Result<Vec<i64>, FormatEr
            }
        )
    {
-        let mut result = vec![0i64; count];
-        // SAFETY: The byte slice is properly aligned for this type and the length is divisible by size_of::<T>().
-        unsafe {
-            core::ptr::copy_nonoverlapping(raw.as_ptr(), result.as_mut_ptr() as *mut u8, raw.len());
-        }
-        return Ok(result);
+        return Ok(native_le_to_vec::<i64>(raw, count));
    }

    let order = get_byte_order(datatype);
@@ -1040,12 +1074,7 @@ pub fn read_as_f32(raw: &[u8], datatype: &Datatype) -> Result<Vec<f32>, FormatEr
            ..
        }
    ) {
-        let mut result = vec![0.0f32; count];
-        // SAFETY: The byte slice is properly aligned for this type and the length is divisible by size_of::<T>().
-        unsafe {
-            core::ptr::copy_nonoverlapping(raw.as_ptr(), result.as_mut_ptr() as *mut u8, raw.len());
-        }
-        return Ok(result);
+        return Ok(native_le_to_vec::<f32>(raw, count));
    }

    let order = get_byte_order(datatype);
@@ -1122,12 +1151,7 @@ pub fn read_as_i32(raw: &[u8], datatype: &Datatype) -> Result<Vec<i32>, FormatEr
            }
        )
    {
-        let mut result = vec![0i32; count];
-        // SAFETY: The byte slice is properly aligned for this type and the length is divisible by size_of::<T>().
-        unsafe {
-            core::ptr::copy_nonoverlapping(raw.as_ptr(), result.as_mut_ptr() as *mut u8, raw.len());
-        }
-        return Ok(result);
+        return Ok(native_le_to_vec::<i32>(raw, count));
    }

    let order = get_byte_order(datatype);
@@ -1403,6 +1427,26 @@ pub fn read_object_references(
            }
            Ok(result)
        }
+        Datatype::Reference {
+            ref_type: crate::datatype::ReferenceType::Object2,
+            size,
+        } => {
+            let elem_size = *size as usize;
+            if elem_size == 0 {
+                return Ok(Vec::new());
+            }
+            if !raw.len().is_multiple_of(elem_size) {
+                return Err(FormatError::DataSizeMismatch {
+                    expected: 0,
+                    actual: raw.len(),
+                });
+            }
+            raw.chunks_exact(elem_size)
+                .map(|element| {
+                    decode_std_object_ref(element).map(|address| ObjectReference { address })
+                })
+                .collect()
+        }
        _ => Err(FormatError::TypeMismatch {
            expected: "Reference(Object)",
            actual: datatype_name(datatype),
@@ -1410,6 +1454,46 @@ pub fn read_object_references(
    }
 }

+/// Decode one `H5T_STD_REF` object reference as stored in a dataset:
+/// `type(1) flags(1) token_size(1) token(token_size)`, zero-padded to the
+/// element size. For a reference within the same file the token is the target
+/// object's header address. An all-zero element is a null reference and
+/// decodes to the undefined address (`u64::MAX`).
+fn decode_std_object_ref(element: &[u8]) -> Result<u64, FormatError> {
+    const STD_REF_OBJECT: u8 = 2;
+    const FLAG_EXTERNAL: u8 = 0x01;
+    if element.iter().all(|&b| b == 0) {
+        return Ok(u64::MAX);
+    }
+    let [ref_type, flags, token_size, token @ ..] = element else {
+        return Err(FormatError::UnexpectedEof {
+            expected: 3,
+            available: element.len(),
+        });
+    };
+    if *ref_type != STD_REF_OBJECT {
+        return Err(FormatError::InvalidReferenceType(*ref_type));
+    }
+    if flags & FLAG_EXTERNAL != 0 {
+        // Carries a file name as well; nothing here follows those.
+        return Err(FormatError::TypeMismatch {
+            expected: "object reference within this file",
+            actual: "external object reference",
+        });
+    }
+    let n = *token_size as usize;
+    if n == 0 || n > 8 || n > token.len() {
+        return Err(FormatError::UnexpectedEof {
+            expected: 3 + n,
+            available: element.len(),
+        });
+    }
+    Ok(token[..n]
+        .iter()
+        .rev()
+        .fold(0u64, |addr, &byte| (addr << 8) | u64::from(byte)))
+}
+
 /// Read region references from raw bytes.
 ///
 /// Region references encode a dataset selection (hyperslab, point list, etc.)
@@ -1763,11 +1847,11 @@ mod tests {
    fn f16_bits(v: f32) -> u16 {
        // Encode a few exact values used by the test.
        match v {
-            x if x == 0.0 => 0x0000,
-            x if x == 1.0 => 0x3c00,
-            x if x == -2.0 => 0xc000,
-            x if x == 0.5 => 0x3800,
-            x if x == 65504.0 => 0x7bff, // f16 max
+            0.0 => 0x0000,
+            1.0 => 0x3c00,
+            -2.0 => 0xc000,
+            0.5 => 0x3800,
+            65504.0 => 0x7bff, // f16 max
            _ => panic!("unsupported test value {v}"),
        }
    }
@@ -2186,7 +2270,7 @@ mod tests {
            ],
        };
        let mut raw = Vec::new();
-        raw.extend_from_slice(&3.14f64.to_le_bytes());
+        raw.extend_from_slice(&3.25f64.to_le_bytes());
        raw.extend_from_slice(&42i32.to_le_bytes());

        let field = read_compound_field(&raw, &dt, "id").unwrap();
@@ -1,5 +1,7 @@
 //! HDF5 Dataspace message parsing (message type 0x0001).

+#[cfg(not(feature = "std"))]
+use alloc::format;
 #[cfg(not(feature = "std"))]
 use alloc::vec::Vec;

@@ -167,6 +169,27 @@ impl Dataspace {
        }
    }

+    /// [`Dataspace::num_elements`] with the product overflow-checked. The
+    /// dimensions are untrusted 64-bit fields; read paths that size a buffer
+    /// from them must use this one.
+    pub fn checked_num_elements(&self) -> Result<u64, FormatError> {
+        match self.space_type {
+            DataspaceType::Null => Ok(0),
+            DataspaceType::Scalar => Ok(1),
+            DataspaceType::Simple if self.dimensions.is_empty() => Ok(0),
+            DataspaceType::Simple => self
+                .dimensions
+                .iter()
+                .try_fold(1u64, |acc, &d| acc.checked_mul(d))
+                .ok_or_else(|| {
+                    FormatError::Overflow(format!(
+                        "dataspace dimensions {:?} overflow the element count",
+                        self.dimensions
+                    ))
+                }),
+        }
+    }
+
    /// Total number of elements. Scalar = 1, Null = 0.
    pub fn num_elements(&self) -> u64 {
        match self.space_type {
@@ -176,7 +199,12 @@ impl Dataspace {
                if self.dimensions.is_empty() {
                    0
                } else {
-                    self.dimensions.iter().product()
+                    // Saturate rather than wrap: a wrapped product could
+                    // under-size a buffer. Size-critical callers use
+                    // `checked_num_elements`.
+                    self.dimensions
+                        .iter()
+                        .fold(1u64, |acc, &d| acc.saturating_mul(d))
                }
            }
        }
@@ -189,11 +217,7 @@ mod tests {

    fn build_v1_dataspace(rank: u8, flags: u8, dims: &[u64], max_dims: Option<&[u64]>) -> Vec<u8> {
        let length_size = 8u8;
-        let mut buf = Vec::new();
-        buf.push(1); // version
-        buf.push(rank);
-        buf.push(flags);
-        buf.push(0); // reserved
+        let mut buf = vec![1, rank, flags, 0]; // version, rank, flags, reserved
        buf.extend_from_slice(&[0u8; 4]); // reserved(4)
        for &d in dims {
            buf.extend_from_slice(&d.to_le_bytes());
@@ -214,11 +238,7 @@ mod tests {
        dims: &[u64],
        max_dims: Option<&[u64]>,
    ) -> Vec<u8> {
-        let mut buf = Vec::new();
-        buf.push(2); // version
-        buf.push(rank);
-        buf.push(flags);
-        buf.push(type_byte);
+        let mut buf = vec![2, rank, flags, type_byte]; // version, rank, flags, type
        for &d in dims {
            buf.extend_from_slice(&d.to_le_bytes());
        }
@@ -298,11 +318,7 @@ mod tests {

    #[test]
    fn v1_with_4byte_length() {
-        let mut buf = Vec::new();
-        buf.push(1); // version
-        buf.push(1); // rank
-        buf.push(0); // flags
-        buf.push(0); // reserved
+        let mut buf = vec![1, 1, 0, 0]; // version, rank, flags, reserved
        buf.extend_from_slice(&[0u8; 4]); // reserved(4)
        buf.extend_from_slice(&10u32.to_le_bytes()); // dim with length_size=4
        let ds = Dataspace::parse(&buf, 4).unwrap();
@@ -36,8 +36,18 @@ pub enum CharacterSet {
 /// Reference type.
 #[derive(Debug, Clone, PartialEq)]
 pub enum ReferenceType {
+    /// Legacy object reference: the target's object header address.
    Object,
+    /// Legacy dataset region reference.
    DatasetRegion,
+    /// `H5T_STD_REF` object reference (HDF5 1.12+, datatype message version
+    /// 4): a small header followed by an object token. Decoded by
+    /// `data_read::read_object_references`.
+    Object2,
+    /// `H5T_STD_REF` dataset region reference.
+    DatasetRegion2,
+    /// `H5T_STD_REF` attribute reference.
+    Attribute,
 }

 /// A member of a compound datatype.
@@ -204,11 +214,25 @@ fn read_uint(data: &[u8], offset: usize, nbytes: usize) -> Result<u64, FormatErr
    })
 }

+/// Maximum recursion depth for nested datatypes (Compound/Enumeration/
+/// VariableLength/Array). A crafted file can nest a message-size-capped
+/// (65535 byte) datatype message ~8000 levels deep, which would blow the
+/// stack — especially on the project's no_std/embedded targets where
+/// available stack is a few KB.
+const MAX_DATATYPE_DEPTH: u16 = 64;
+
 impl Datatype {
    /// Parse a datatype message from raw bytes.
    ///
    /// Returns `(Datatype, bytes_consumed)` for recursive parsing.
    pub fn parse(data: &[u8]) -> Result<(Datatype, usize), FormatError> {
+        Self::parse_with_depth(data, 0)
+    }
+
+    fn parse_with_depth(data: &[u8], depth: u16) -> Result<(Datatype, usize), FormatError> {
+        if depth >= MAX_DATATYPE_DEPTH {
+            return Err(FormatError::NestingDepthExceeded);
+        }
        // Minimum header: 4 bytes (class_and_version + 3 bytes bit field) + 4 bytes size = 8
        ensure_len(data, 0, 8)?;

@@ -358,7 +382,8 @@ impl Datatype {
                        pos += name_len;
                        let byte_offset = read_uint(data, pos, ob)?;
                        pos += ob;
-                        let (member_dt, consumed) = Datatype::parse(&data[pos..])?;
+                        let (member_dt, consumed) =
+                            Self::parse_with_depth(&data[pos..], depth + 1)?;
                        pos += consumed;
                        members.push(CompoundMember {
                            name,
@@ -367,24 +392,29 @@ impl Datatype {
                        });
                    }
                } else if version == 1 || version == 2 {
-                    // v1/v2: name, offset(4), dimensionality(1), reserved(3), dim_perm(4),
-                    //         reserved_dims(up to 4*4=16), member datatype
+                    // v1/v2: name (null-terminated, padded to a multiple of 8
+                    // bytes), offset(4), member datatype. v1 additionally
+                    // carries the legacy per-member array fields between the
+                    // offset and the member datatype: dimensionality(1),
+                    // reserved(3), dim_perm(4), reserved(4), 4 dim sizes(16).
+                    // v1 is what default (non-`latest`) libver bounds emit.
                    for _ in 0..num_members {
                        let (name, name_len) = read_null_terminated_string(data, pos)?;
-                        pos += name_len;
-                        // v1: names padded to 8-byte boundary
-                        if version == 1 {
-                            let total_name_bytes = name_len;
-                            let padded = (total_name_bytes + 7) & !7;
-                            pos = pos - name_len + padded;
-                        }
+                        let padded = name_len.checked_add(7).ok_or(FormatError::UnexpectedEof {
+                            expected: usize::MAX,
+                            available: data.len(),
+                        })? & !7;
+                        ensure_len(data, pos, padded)?;
+                        pos += padded;
                        ensure_len(data, pos, 4)?;
                        let byte_offset = LittleEndian::read_u32(&data[pos..pos + 4]) as u64;
                        pos += 4;
-                        // dimensionality(1) + reserved(3) + dim_perm(4) + 4 dim slots(16) = 24
-                        ensure_len(data, pos, 24)?;
-                        pos += 24;
-                        let (member_dt, consumed) = Datatype::parse(&data[pos..])?;
+                        if version == 1 {
+                            ensure_len(data, pos, 28)?;
+                            pos += 28;
+                        }
+                        let (member_dt, consumed) =
+                            Self::parse_with_depth(&data[pos..], depth + 1)?;
                        pos += consumed;
                        members.push(CompoundMember {
                            name,
@@ -404,9 +434,15 @@ impl Datatype {
            7 => {
                // Reference
                let ref_type_val = bf0 & 0x0F;
-                let ref_type = match ref_type_val {
-                    0 => ReferenceType::Object,
-                    1 => ReferenceType::DatasetRegion,
+                // Datatype message version 4 (HDF5 1.12) revised this class:
+                // types 2-4 are the new `H5T_STD_REF` references, and the high
+                // nibble of the first flag byte carries their encoding version.
+                let ref_type = match (ref_type_val, version) {
+                    (0, _) => ReferenceType::Object,
+                    (1, _) => ReferenceType::DatasetRegion,
+                    (2, 4..) => ReferenceType::Object2,
+                    (3, 4..) => ReferenceType::DatasetRegion2,
+                    (4, 4..) => ReferenceType::Attribute,
                    _ => return Err(FormatError::InvalidReferenceType(ref_type_val)),
                };
                Ok((Datatype::Reference { size, ref_type }, pos))
@@ -415,7 +451,7 @@ impl Datatype {
                // Enumeration
                let num_members = (bf0 as u16) | ((bf1 as u16) << 8);
                // Parse base type
-                let (base_type, base_consumed) = Datatype::parse(&data[pos..])?;
+                let (base_type, base_consumed) = Self::parse_with_depth(&data[pos..], depth + 1)?;
                pos += base_consumed;
                let base_size = base_type.type_size();
                let mut members = Vec::with_capacity(num_members as usize);
@@ -468,7 +504,7 @@ impl Datatype {
                } else {
                    None
                };
-                let (base_type, consumed) = Datatype::parse(&data[pos..])?;
+                let (base_type, consumed) = Self::parse_with_depth(&data[pos..], depth + 1)?;
                pos += consumed;
                Ok((
                    Datatype::VariableLength {
@@ -494,7 +530,7 @@ impl Datatype {
                    }
                    // skip permutation indices
                    pos += ndims * 4;
-                    let (base_type, consumed) = Datatype::parse(&data[pos..])?;
+                    let (base_type, consumed) = Self::parse_with_depth(&data[pos..], depth + 1)?;
                    pos += consumed;
                    Ok((
                        Datatype::Array {
@@ -515,7 +551,7 @@ impl Datatype {
                        dimensions.push(LittleEndian::read_u32(&data[pos..pos + 4]));
                        pos += 4;
                    }
-                    let (base_type, consumed) = Datatype::parse(&data[pos..])?;
+                    let (base_type, consumed) = Self::parse_with_depth(&data[pos..], depth + 1)?;
                    pos += consumed;
                    Ok((
                        Datatype::Array {
@@ -532,27 +568,39 @@ impl Datatype {
                }
            }
            11 => {
-                // Complex number — store as compound of two floats internally
-                // Parse like compound with version 3 and 2 members
-                // But actually class 11 has no special properties beyond class 6 compound.
-                // It's just recognized as a separate class. For now parse the 2 members
-                // as compound.
-                let num_members = (bf0 as u16) | ((bf1 as u16) << 8);
-                let mut members = Vec::with_capacity(num_members as usize);
-                let ob = offset_bytes_for_size(size);
-                for _ in 0..num_members {
-                    let (name, name_len) = read_null_terminated_string(data, pos)?;
-                    pos += name_len;
-                    let byte_offset = read_uint(data, pos, ob)?;
-                    pos += ob;
-                    let (member_dt, consumed) = Datatype::parse(&data[pos..])?;
-                    pos += consumed;
-                    members.push(CompoundMember {
-                        name,
-                        byte_offset,
-                        datatype: member_dt,
+                // Complex number (HDF5 2.0, datatype version 5). The properties
+                // are a single base floating-point datatype message; an element
+                // is two consecutive base-type values (real, imaginary). There
+                // is no member list. Surface it as the equivalent two-member
+                // compound `{r, i}` — the same shape h5py writes for numpy
+                // complex dtypes — so downstream compound readers work as-is.
+                if version != 5 {
+                    return Err(FormatError::InvalidDatatypeVersion {
+                        class: class_id,
+                        version,
                    });
                }
+                let (base_type, consumed) = Self::parse_with_depth(&data[pos..], depth + 1)?;
+                pos += consumed;
+                let base_size = base_type.type_size();
+                if base_size.checked_mul(2) != Some(size) {
+                    return Err(FormatError::DataSizeMismatch {
+                        expected: (base_size as usize).saturating_mul(2),
+                        actual: size as usize,
+                    });
+                }
+                let members = vec![
+                    CompoundMember {
+                        name: String::from("r"),
+                        byte_offset: 0,
+                        datatype: base_type.clone(),
+                    },
+                    CompoundMember {
+                        name: String::from("i"),
+                        byte_offset: base_size as u64,
+                        datatype: base_type,
+                    },
+                ];
                Ok((Datatype::Compound { size, members }, pos))
            }
            _ => Err(FormatError::InvalidDatatypeClass(class_id)),
@@ -814,6 +862,39 @@ mod tests {
        buf
    }

+    /// A crafted datatype message nesting Variable-Length wrappers deeper
+    /// than `MAX_DATATYPE_DEPTH` must return `NestingDepthExceeded`
+    /// instead of overflowing the stack.
+    #[test]
+    fn nested_variable_length_exceeds_depth_limit() {
+        // Each VL level is just an 8-byte header (class 9, vl_type=0 =>
+        // sequence, no padding/charset fields) immediately followed by the
+        // next level's bytes, terminated by a fixed-point base type.
+        let levels = MAX_DATATYPE_DEPTH as usize + 10;
+        let mut data = Vec::new();
+        for _ in 0..levels {
+            data.extend_from_slice(&build_dt_header(9, 3, [0, 0, 0], 0));
+        }
+        data.extend_from_slice(&build_fixed_point(4, false, false, 0, 32));
+
+        let result = Datatype::parse(&data);
+        assert!(matches!(result, Err(FormatError::NestingDepthExceeded)));
+    }
+
+    /// A datatype nested just within the depth limit must still parse fine.
+    #[test]
+    fn nested_variable_length_within_depth_limit_ok() {
+        let levels = MAX_DATATYPE_DEPTH as usize - 1;
+        let mut data = Vec::new();
+        for _ in 0..levels {
+            data.extend_from_slice(&build_dt_header(9, 3, [0, 0, 0], 0));
+        }
+        data.extend_from_slice(&build_fixed_point(4, false, false, 0, 32));
+
+        let result = Datatype::parse(&data);
+        assert!(result.is_ok());
+    }
+
    #[test]
    fn test_fixed_point_u8() {
        let data = build_fixed_point(1, false, false, 0, 8);
@@ -1079,6 +1160,156 @@ mod tests {
        }
    }

+    /// Real datatype message bytes emitted by h5py 3.16 / HDF5 2.0 with
+    /// *default* libver bounds for [('x','f8'),('y','f8'),('id','i4')]:
+    /// compound datatype version 1 (padded names + 28 bytes of legacy
+    /// per-member array fields).
+    fn compound_v1_bytes() -> Vec<u8> {
+        let f64le: [u8; 20] = [
+            0x11, 0x20, 0x3f, 0x00, 0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x40, 0x00, 0x34, 0x0b,
+            0x00, 0x34, 0xff, 0x03, 0x00, 0x00,
+        ];
+        let i32le: [u8; 12] = [
+            0x10, 0x08, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x20, 0x00,
+        ];
+        let mut b = vec![0x16, 0x03, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00];
+        for (name, offset, dt) in [
+            (&b"x"[..], 0u32, &f64le[..]),
+            (&b"y"[..], 8, &f64le[..]),
+            (&b"id"[..], 16, &i32le[..]),
+        ] {
+            let mut padded = name.to_vec();
+            padded.resize((name.len() + 1 + 7) & !7, 0);
+            b.extend_from_slice(&padded);
+            b.extend_from_slice(&offset.to_le_bytes());
+            b.extend_from_slice(&[0u8; 28]);
+            b.extend_from_slice(dt);
+        }
+        b
+    }
+
+    fn assert_xyid_compound(dt: Datatype) {
+        match dt {
+            Datatype::Compound { size, members } => {
+                assert_eq!(size, 20);
+                let got: Vec<(&str, u64, u32)> = members
+                    .iter()
+                    .map(|m| (m.name.as_str(), m.byte_offset, m.datatype.type_size()))
+                    .collect();
+                assert_eq!(got, vec![("x", 0, 8), ("y", 8, 8), ("id", 16, 4)]);
+            }
+            other => panic!("expected Compound, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn test_compound_v1_default_libver() {
+        let bytes = compound_v1_bytes();
+        let (dt, consumed) = Datatype::parse(&bytes).unwrap();
+        assert_eq!(consumed, bytes.len());
+        assert_xyid_compound(dt);
+    }
+
+    #[test]
+    fn test_compound_v2_padded_names_no_array_fields() {
+        // v2 = v1 without the 28 bytes of per-member array fields; names are
+        // still padded to a multiple of 8 (matches libhdf5's H5O decoder).
+        let v1 = compound_v1_bytes();
+        let mut v2 = vec![0x26, 0x03, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00];
+        let mut pos = 8;
+        for dt_len in [20usize, 20, 12] {
+            v2.extend_from_slice(&v1[pos..pos + 8 + 4]); // padded name + offset
+            pos += 8 + 4 + 28;
+            v2.extend_from_slice(&v1[pos..pos + dt_len]);
+            pos += dt_len;
+        }
+        let (dt, consumed) = Datatype::parse(&v2).unwrap();
+        assert_eq!(consumed, v2.len());
+        assert_xyid_compound(dt);
+    }
+
+    #[test]
+    fn test_compound_v1_truncated_is_error_not_panic() {
+        let bytes = compound_v1_bytes();
+        for cut in 8..bytes.len() {
+            assert!(Datatype::parse(&bytes[..cut]).is_err(), "cut at {cut}");
+        }
+    }
+
+    /// Real datatype message bytes emitted by HDF5 2.0 for the native complex
+    /// type `H5T_COMPLEX_IEEE_F64LE`: class 11, version 5, size 16, followed by
+    /// the base IEEE f64 datatype message.
+    const COMPLEX_F64_HDF5_2_0: [u8; 28] = [
+        0x5b, 0x01, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x11, 0x20, 0x3f, 0x00, 0x08, 0x00, 0x00,
+        0x00, 0x00, 0x00, 0x40, 0x00, 0x34, 0x0b, 0x00, 0x34, 0xff, 0x03, 0x00, 0x00,
+    ];
+
+    #[test]
+    fn test_complex_v5_from_hdf5_2_0() {
+        let (dt, consumed) = Datatype::parse(&COMPLEX_F64_HDF5_2_0).unwrap();
+        assert_eq!(consumed, COMPLEX_F64_HDF5_2_0.len());
+        match dt {
+            Datatype::Compound { size, members } => {
+                assert_eq!(size, 16);
+                assert_eq!(members.len(), 2);
+                assert_eq!((members[0].name.as_str(), members[0].byte_offset), ("r", 0));
+                assert_eq!((members[1].name.as_str(), members[1].byte_offset), ("i", 8));
+                for m in &members {
+                    assert!(matches!(
+                        m.datatype,
+                        Datatype::FloatingPoint { size: 8, .. }
+                    ));
+                }
+            }
+            other => panic!("expected Compound, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn test_compound_with_complex_member_from_hdf5_2_0() {
+        // Compound { z: complex f64 @0, k: i64 @16 } as written by HDF5 2.0.
+        // Regression guard: the complex member must consume exactly its own
+        // bytes so the following member parses.
+        let mut bytes = vec![
+            0x56, 0x02, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, b'z', 0x00, 0x00,
+        ];
+        bytes.extend_from_slice(&COMPLEX_F64_HDF5_2_0);
+        bytes.extend_from_slice(&[b'k', 0x00, 0x10]);
+        bytes.extend_from_slice(&[
+            0x10, 0x08, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x40, 0x00,
+        ]);
+        let (dt, consumed) = Datatype::parse(&bytes).unwrap();
+        assert_eq!(consumed, bytes.len());
+        match dt {
+            Datatype::Compound { size, members } => {
+                assert_eq!(size, 24);
+                assert_eq!(members.len(), 2);
+                assert!(matches!(
+                    &members[0].datatype,
+                    Datatype::Compound { size: 16, members } if members.len() == 2
+                ));
+                assert_eq!(
+                    (members[1].name.as_str(), members[1].byte_offset),
+                    ("k", 16)
+                );
+            }
+            other => panic!("expected Compound, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn test_complex_size_mismatch_rejected() {
+        let mut bytes = COMPLEX_F64_HDF5_2_0;
+        bytes[4] = 0x0c; // claims 12 bytes, base type is 8
+        assert!(matches!(
+            Datatype::parse(&bytes),
+            Err(FormatError::DataSizeMismatch {
+                expected: 16,
+                actual: 12
+            })
+        ));
+    }
+
    #[test]
    fn test_reference_object() {
        let buf = build_dt_header(7, 1, [0, 0, 0], 8);
@@ -1348,6 +1579,28 @@ mod tests {
        assert_eq!(err, FormatError::InvalidCharacterSet(2));
    }

+    #[test]
+    fn test_reference_v4_std_ref_from_hdf5_2_0() {
+        // Datatype message of an H5T_STD_REF dataset written by HDF5 2.0:
+        // class 7, version 4, type 2 (object), encoding version 1, 18 bytes.
+        let bytes = [0x47, 0x12, 0x00, 0x00, 0x12, 0x00, 0x00, 0x00];
+        let (dt, consumed) = Datatype::parse(&bytes).unwrap();
+        assert_eq!(consumed, 8);
+        assert_eq!(
+            dt,
+            Datatype::Reference {
+                size: 18,
+                ref_type: ReferenceType::Object2
+            }
+        );
+        // The new types are only valid from datatype version 4.
+        let old_version = [0x37, 0x12, 0x00, 0x00, 0x12, 0x00, 0x00, 0x00];
+        assert_eq!(
+            Datatype::parse(&old_version).unwrap_err(),
+            FormatError::InvalidReferenceType(2)
+        );
+    }
+
    #[test]
    fn test_error_invalid_reference_type() {
        let buf = build_dt_header(7, 1, [5, 0, 0], 8);
@@ -114,6 +114,23 @@ pub enum FormatError {
    InvalidAttributeInfoVersion(u8),
    /// Invalid shared message version.
    InvalidSharedMessageVersion(u8),
+    /// A message is marked shared but was parsed without access to the file,
+    /// so the reference to the real message could not be followed.
+    UnresolvedSharedMessage,
+    /// A selection does not fit the dataset it was applied to (wrong rank, or
+    /// it reaches past a dimension's extent).
+    SelectionOutOfBounds(String),
+    /// The dataset's raw data is stored in external files (External Data
+    /// Files message), which this reader does not follow.
+    ExternalDataFilesUnsupported,
+    /// The path goes through an external link (a link into another file),
+    /// which this reader does not follow.
+    ExternalLinkUnsupported {
+        /// The file the link points into.
+        filename: String,
+        /// The object path within that file.
+        object_path: String,
+    },
    /// Invalid SOHM table version.
    InvalidSohmTableVersion(u8),
    /// Invalid SOHM table signature (expected "SMTB").
@@ -307,6 +324,25 @@ impl fmt::Display for FormatError {
            FormatError::InvalidSharedMessageVersion(v) => {
                write!(f, "invalid shared message version: {v}")
            }
+            FormatError::ExternalLinkUnsupported {
+                filename,
+                object_path,
+            } => write!(
+                f,
+                "path goes through an external link to {object_path} in {filename}, which is \
+                 not supported"
+            ),
+            FormatError::ExternalDataFilesUnsupported => write!(
+                f,
+                "dataset raw data is stored in external file(s), which is not supported"
+            ),
+            FormatError::SelectionOutOfBounds(msg) => {
+                write!(f, "selection out of bounds: {msg}")
+            }
+            FormatError::UnresolvedSharedMessage => write!(
+                f,
+                "message is shared but no file data was available to resolve it"
+            ),
            FormatError::InvalidSohmTableVersion(v) => {
                write!(f, "invalid SOHM table version: {v}")
            }
@@ -54,6 +54,19 @@ fn read_offset(data: &[u8], pos: usize, size: u8) -> Result<u64, FormatError> {
    })
 }

+fn ensure_len(data: &[u8], offset: usize, needed: usize) -> Result<(), FormatError> {
+    if offset
+        .checked_add(needed)
+        .is_none_or(|end| end > data.len())
+    {
+        return Err(FormatError::UnexpectedEof {
+            expected: offset.saturating_add(needed),
+            available: data.len(),
+        });
+    }
+    Ok(())
+}
+
 fn is_undefined_addr(addr: u64, offset_size: u8) -> bool {
    match offset_size {
        2 => addr == 0xFFFF,
@@ -98,12 +111,7 @@ impl ExtensibleArrayHeader {
        //   6 stats fields (each length_size) + index_block_address(offset_size) + checksum(4)
        let min_size =
            4 + 1 + 1 + 1 + 1 + 1 + 1 + 1 + 1 + 6 * length_size as usize + offset_size as usize + 4;
-        if offset + min_size > file_data.len() {
-            return Err(FormatError::UnexpectedEof {
-                expected: offset + min_size,
-                available: file_data.len(),
-            });
-        }
+        ensure_len(file_data, offset, min_size)?;

        let d = &file_data[offset..];
        if &d[0..4] != b"EAHD" {
@@ -275,12 +283,7 @@ fn read_data_block_elements(
 ) -> Result<Vec<ChunkInfo>, FormatError> {
    // AEDB: signature(4) + version(1) + client_id(1) + header_address(offset_size)
    let db_header_size = 4 + 1 + 1 + offset_size as usize;
-    if db_offset + db_header_size > file_data.len() {
-        return Err(FormatError::UnexpectedEof {
-            expected: db_offset + db_header_size,
-            available: file_data.len(),
-        });
-    }
+    ensure_len(file_data, db_offset, db_header_size)?;

    let d = &file_data[db_offset..];
    if &d[0..4] != b"EADB" {
@@ -427,12 +430,7 @@ pub fn read_extensible_array_chunks(
    // Parse index block (AEIB)
    let ib_offset = header.index_block_address as usize;
    let ib_header_size = 4 + 1 + 1 + offset_size as usize; // sig + ver + client + hdr_addr
-    if ib_offset + ib_header_size > file_data.len() {
-        return Err(FormatError::UnexpectedEof {
-            expected: ib_offset + ib_header_size,
-            available: file_data.len(),
-        });
-    }
+    ensure_len(file_data, ib_offset, ib_header_size)?;

    let ib = &file_data[ib_offset..];
    if &ib[0..4] != b"EAIB" {
@@ -628,12 +626,7 @@ fn read_super_block(

    // AESB: signature(4) + version(1) + client_id(1) + header_address(offset_size)
    let sb_header_size = 4 + 1 + 1 + os;
-    if sb_offset + sb_header_size > file_data.len() {
-        return Err(FormatError::UnexpectedEof {
-            expected: sb_offset + sb_header_size,
-            available: file_data.len(),
-        });
-    }
+    ensure_len(file_data, sb_offset, sb_header_size)?;

    if &file_data[sb_offset..sb_offset + 4] != b"EASB" {
        return Err(FormatError::ChunkedReadError(
@@ -759,6 +752,33 @@ mod tests {
        assert!(result.is_err());
    }

+    /// A near-`usize::MAX` offset must error cleanly, not overflow/panic.
+    #[test]
+    fn parse_rejects_offset_overflow() {
+        let buf = vec![0u8; 64];
+        let result = ExtensibleArrayHeader::parse(&buf, usize::MAX - 4, 8, 8);
+        assert!(result.is_err());
+    }
+
+    /// A near-`usize::MAX` index block address must error cleanly, not overflow/panic.
+    #[test]
+    fn read_rejects_index_block_offset_overflow() {
+        let header = ExtensibleArrayHeader {
+            client_id: 0,
+            element_size: 8,
+            max_nelmts_bits: 10,
+            idx_blk_elmts: 2,
+            min_dblk_nelmts: 4,
+            super_blk_min_nelmts: 2,
+            max_dblk_nelmts_bits: 8,
+            num_elements: 5,
+            index_block_address: (usize::MAX - 4) as u64,
+        };
+        let buf = vec![0u8; 64];
+        let r = read_extensible_array_chunks(&buf, &header, &[100], &[20], 8, 8, 8);
+        assert!(r.is_err());
+    }
+
    #[test]
    fn parse_header_invalid_version() {
        let mut buf = vec![0u8; 256];
@@ -1221,8 +1221,10 @@ impl FileWriter {
                    precompressed: None,
                });
            } else if is_chunked[i] {
-                let chunk_dims = d.chunk_options.resolve_chunk_dims(&d.ds.dimensions);
                let elem_size = d.dt.type_size() as usize;
+                let chunk_dims = d
+                    .chunk_options
+                    .resolve_chunk_dims_for(&d.ds.dimensions, elem_size);
                // Compress once in Pass 1; cache the result so Pass 2 can skip
                // re-compression and just rebuild the index with real addresses.
                let pre = precompress_chunks(
@@ -0,0 +1,407 @@
+//! Fill Value messages (0x0005, and the old 0x0004) and applying them on read.
+//!
+//! HDF5 allocates storage lazily: a chunk nobody wrote to does not exist in the
+//! file, and a contiguous dataset nobody wrote to has no data address at all.
+//! Reading such a region must yield the dataset's *fill value* (zeros unless
+//! the creator chose otherwise). The readers in [`crate::chunked_read`] leave
+//! those regions zeroed; [`apply_to_unallocated_chunks`] then overwrites exactly
+//! the chunk-grid cells that are absent from the chunk index — so it can never
+//! mistake a stored zero for a hole — and is skipped entirely in the common
+//! case of a zero fill value.
+
+#[cfg(not(feature = "std"))]
+use alloc::{format, vec, vec::Vec};
+
+use crate::chunked_read::{alloc_output, checked_byte_len, list_chunks};
+use crate::data_layout::DataLayout;
+use crate::dataspace::Dataspace;
+use crate::error::FormatError;
+use crate::message_type::MessageType;
+use crate::object_header::HeaderMessage;
+
+/// Largest fill value accepted. A fill value is one element of the dataset's
+/// datatype; this only bounds the allocation driven by the message's size field.
+const MAX_FILL_VALUE_SIZE: usize = 1 << 20;
+
+/// Parse a Fill Value message, returning the user-defined fill value bytes, or
+/// `None` when the dataset uses the default (all zeros) or has the fill value
+/// explicitly undefined.
+pub fn parse_fill_value(msg: &HeaderMessage) -> Result<Option<Vec<u8>>, FormatError> {
+    let data = msg.data.as_slice();
+    let value_at = |pos: usize| -> Result<Option<Vec<u8>>, FormatError> {
+        let size_bytes = data.get(pos..pos + 4).ok_or(FormatError::UnexpectedEof {
+            expected: pos + 4,
+            available: data.len(),
+        })?;
+        let size = u32::from_le_bytes([size_bytes[0], size_bytes[1], size_bytes[2], size_bytes[3]])
+            as usize;
+        if size == 0 {
+            return Ok(None);
+        }
+        if size > MAX_FILL_VALUE_SIZE {
+            return Err(FormatError::Overflow(format!(
+                "fill value of {size} bytes exceeds the {MAX_FILL_VALUE_SIZE}-byte limit"
+            )));
+        }
+        let start = pos + 4;
+        let value =
+            data.get(start..start.saturating_add(size))
+                .ok_or(FormatError::UnexpectedEof {
+                    expected: start.saturating_add(size),
+                    available: data.len(),
+                })?;
+        Ok(Some(value.to_vec()))
+    };
+
+    match msg.msg_type {
+        // Old fill value message: size(4), value.
+        MessageType::FillValueOld => value_at(0),
+        MessageType::FillValue => {
+            let version = *data.first().ok_or(FormatError::UnexpectedEof {
+                expected: 1,
+                available: 0,
+            })?;
+            match version {
+                // version, alloc time, write time, defined, [size, value]
+                1 | 2 => {
+                    let defined = *data.get(3).ok_or(FormatError::UnexpectedEof {
+                        expected: 4,
+                        available: data.len(),
+                    })?;
+                    if version == 2 && defined == 0 {
+                        Ok(None)
+                    } else if data.len() < 8 && version == 1 {
+                        // v1 always carries a size, but tolerate its absence.
+                        Ok(None)
+                    } else {
+                        value_at(4)
+                    }
+                }
+                // version, flags (bit 4 = undefined, bit 5 = defined), [size, value]
+                3 => {
+                    let flags = *data.get(1).ok_or(FormatError::UnexpectedEof {
+                        expected: 2,
+                        available: data.len(),
+                    })?;
+                    if flags & 0x10 != 0 || flags & 0x20 == 0 {
+                        Ok(None)
+                    } else {
+                        value_at(2)
+                    }
+                }
+                v => Err(FormatError::UnsupportedVersion(v)),
+            }
+        }
+        _ => Ok(None),
+    }
+}
+
+/// The fill value that applies to a dataset given its header messages. The new
+/// message wins over the old one when both are present.
+pub fn dataset_fill_value(messages: &[HeaderMessage]) -> Result<Option<Vec<u8>>, FormatError> {
+    for wanted in [MessageType::FillValue, MessageType::FillValueOld] {
+        if let Some(msg) = messages.iter().find(|m| m.msg_type == wanted) {
+            if crate::shared_message::is_shared(msg.flags) {
+                // A shared fill value is legal but vanishingly rare; treat it
+                // as the default rather than misparsing the reference.
+                return Ok(None);
+            }
+            if let Some(value) = parse_fill_value(msg)? {
+                return Ok(Some(value));
+            }
+        }
+    }
+    Ok(None)
+}
+
+/// `true` when a fill value is absent or all zeros, i.e. identical to what the
+/// readers already produce for unallocated storage.
+pub fn is_default(fill: Option<&[u8]>) -> bool {
+    fill.is_none_or(|f| f.iter().all(|&b| b == 0))
+}
+
+/// A whole dataset's worth of fill value: what reading a dataset with no
+/// allocated storage at all must return.
+pub fn filled_dataset(
+    dataspace: &Dataspace,
+    elem_size: usize,
+    fill: Option<&[u8]>,
+) -> Result<Vec<u8>, FormatError> {
+    let total = checked_byte_len(dataspace.checked_num_elements()?, elem_size)?;
+    let mut out = alloc_output(total)?;
+    if let Some(fill) = fill.filter(|f| f.len() == elem_size && !is_default(Some(f))) {
+        for element in out.chunks_exact_mut(elem_size) {
+            element.copy_from_slice(fill);
+        }
+    }
+    Ok(out)
+}
+
+/// Whether the layout has any storage in the file at all. A dataset that was
+/// created but never written to has none.
+pub fn has_storage(layout: &DataLayout) -> bool {
+    !matches!(
+        layout,
+        DataLayout::Contiguous { address: None, .. }
+            | DataLayout::Chunked {
+                btree_address: None,
+                ..
+            }
+    )
+}
+
+/// Run a full-dataset `read`, giving unallocated storage its fill value: a
+/// dataset with no storage at all reads as entirely fill value (instead of
+/// failing), and a chunked dataset has the fill value written into every
+/// chunk the file never allocated.
+#[allow(clippy::too_many_arguments)]
+pub fn read_full_with_fill<E: From<FormatError>>(
+    messages: &[HeaderMessage],
+    file_data: &[u8],
+    layout: &DataLayout,
+    dataspace: &Dataspace,
+    elem_size: usize,
+    offset_size: u8,
+    length_size: u8,
+    read: impl FnOnce() -> Result<Vec<u8>, E>,
+) -> Result<Vec<u8>, E> {
+    // A dataset with external raw data also has no data address in this
+    // file. It is NOT unallocated — its values live elsewhere — so it must
+    // never be answered with the fill value.
+    if messages
+        .iter()
+        .any(|m| m.msg_type == MessageType::ExternalDataFiles)
+    {
+        return Err(FormatError::ExternalDataFilesUnsupported.into());
+    }
+    let fill = dataset_fill_value(messages)?;
+    if !has_storage(layout) {
+        return Ok(filled_dataset(dataspace, elem_size, fill.as_deref())?);
+    }
+    let mut output = read()?;
+    apply_to_unallocated_chunks(
+        &mut output,
+        file_data,
+        layout,
+        dataspace,
+        elem_size,
+        fill.as_deref(),
+        offset_size,
+        length_size,
+    )?;
+    Ok(output)
+}
+
+/// Overwrite, in a fully read chunked dataset `output`, every region whose
+/// chunk was never allocated with `fill`. No-op for non-chunked layouts, a
+/// default fill value, or a fill value whose size doesn't match the element.
+#[allow(clippy::too_many_arguments)]
+pub fn apply_to_unallocated_chunks(
+    output: &mut [u8],
+    file_data: &[u8],
+    layout: &DataLayout,
+    dataspace: &Dataspace,
+    elem_size: usize,
+    fill: Option<&[u8]>,
+    offset_size: u8,
+    length_size: u8,
+) -> Result<(), FormatError> {
+    let Some(fill) = fill.filter(|f| f.len() == elem_size && !is_default(Some(f))) else {
+        return Ok(());
+    };
+    if !matches!(layout, DataLayout::Chunked { .. }) || elem_size == 0 {
+        return Ok(());
+    }
+    let (chunks, chunk_dims) = list_chunks(
+        file_data,
+        layout,
+        dataspace,
+        elem_size,
+        offset_size,
+        length_size,
+    )?;
+    let rank = chunk_dims.len();
+    let ds_dims: Vec<usize> = dataspace.dimensions.iter().map(|&d| d as usize).collect();
+    if rank == 0 || ds_dims.len() != rank || chunk_dims.contains(&0) {
+        return Ok(());
+    }
+
+    // Row-major strides over the dataset and over the chunk grid.
+    let mut ds_strides = vec![1usize; rank];
+    for i in (0..rank - 1).rev() {
+        ds_strides[i] = ds_strides[i + 1].saturating_mul(ds_dims[i + 1]);
+    }
+    let grid: Vec<usize> = ds_dims
+        .iter()
+        .zip(&chunk_dims)
+        .map(|(&d, &c)| d.div_ceil(c))
+        .collect();
+    let cells = grid
+        .iter()
+        .try_fold(1usize, |acc, &g| acc.checked_mul(g))
+        .ok_or_else(|| FormatError::Overflow("chunk grid size overflows".into()))?;
+    if cells == 0 {
+        return Ok(());
+    }
+
+    let mut allocated = vec![false; cells];
+    for chunk in &chunks {
+        // Undefined address: the index has a slot for the chunk but no storage.
+        if chunk.address == u64::MAX || chunk.offsets.len() < rank {
+            continue;
+        }
+        let mut cell = 0usize;
+        let mut in_range = true;
+        for d in 0..rank {
+            let coord = chunk.offsets[d] as usize / chunk_dims[d];
+            if coord >= grid[d] {
+                in_range = false;
+                break;
+            }
+            cell = cell * grid[d] + coord;
+        }
+        if in_range {
+            allocated[cell] = true;
+        }
+    }
+
+    let mut coord = vec![0usize; rank];
+    for (cell, is_allocated) in allocated.iter().enumerate() {
+        if *is_allocated {
+            continue;
+        }
+        // Decode the cell index into grid coordinates.
+        let mut rem = cell;
+        for d in (0..rank).rev() {
+            coord[d] = rem % grid[d];
+            rem /= grid[d];
+        }
+        fill_cell(
+            output,
+            &coord,
+            &chunk_dims,
+            &ds_dims,
+            &ds_strides,
+            elem_size,
+            fill,
+        );
+    }
+    Ok(())
+}
+
+/// Fill the part of chunk-grid cell `coord` that lies inside the dataset.
+fn fill_cell(
+    output: &mut [u8],
+    coord: &[usize],
+    chunk_dims: &[usize],
+    ds_dims: &[usize],
+    ds_strides: &[usize],
+    elem_size: usize,
+    fill: &[u8],
+) {
+    let rank = coord.len();
+    let start: Vec<usize> = (0..rank).map(|d| coord[d] * chunk_dims[d]).collect();
+    let end: Vec<usize> = (0..rank)
+        .map(|d| (start[d] + chunk_dims[d]).min(ds_dims[d]))
+        .collect();
+    if (0..rank).any(|d| start[d] >= end[d]) {
+        return;
+    }
+    // Walk every row (all dims but the last) and fill the run along the last.
+    let run = end[rank - 1] - start[rank - 1];
+    let mut idx = start.clone();
+    loop {
+        let first: usize = (0..rank).map(|d| idx[d] * ds_strides[d]).sum();
+        let from = first * elem_size;
+        let to = from + run * elem_size;
+        if let Some(region) = output.get_mut(from..to) {
+            for element in region.chunks_exact_mut(elem_size) {
+                element.copy_from_slice(fill);
+            }
+        }
+        // Advance the odometer over dims 0..rank-1.
+        let mut d = rank - 1;
+        loop {
+            if d == 0 {
+                return;
+            }
+            d -= 1;
+            idx[d] += 1;
+            if idx[d] < end[d] {
+                break;
+            }
+            idx[d] = start[d];
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn msg(msg_type: MessageType, data: &[u8]) -> HeaderMessage {
+        HeaderMessage {
+            msg_type,
+            size: data.len(),
+            flags: 0,
+            creation_order: None,
+            data: data.to_vec(),
+        }
+    }
+
+    #[test]
+    fn parses_v3_defined_undefined_and_default() {
+        // Real message for h5py `fillvalue=-1` on an i4 dataset (HDF5 2.0).
+        let defined = msg(
+            MessageType::FillValue,
+            &[3, 0x2b, 4, 0, 0, 0, 0xff, 0xff, 0xff, 0xff],
+        );
+        assert_eq!(parse_fill_value(&defined).unwrap(), Some(vec![0xff; 4]));
+        let default = msg(MessageType::FillValue, &[3, 0x0a]);
+        assert_eq!(parse_fill_value(&default).unwrap(), None);
+        let undefined = msg(MessageType::FillValue, &[3, 0x19]);
+        assert_eq!(parse_fill_value(&undefined).unwrap(), None);
+    }
+
+    #[test]
+    fn parses_v2_and_old_messages() {
+        let v2 = msg(MessageType::FillValue, &[2, 2, 2, 1, 2, 0, 0, 0, 7, 0]);
+        assert_eq!(parse_fill_value(&v2).unwrap(), Some(vec![7, 0]));
+        let v2_undefined = msg(MessageType::FillValue, &[2, 2, 2, 0]);
+        assert_eq!(parse_fill_value(&v2_undefined).unwrap(), None);
+        let old = msg(MessageType::FillValueOld, &[2, 0, 0, 0, 9, 9]);
+        assert_eq!(parse_fill_value(&old).unwrap(), Some(vec![9, 9]));
+    }
+
+    #[test]
+    fn truncated_or_oversized_fill_is_an_error() {
+        let short = msg(MessageType::FillValue, &[3, 0x29, 4, 0, 0, 0, 0xff]);
+        assert!(parse_fill_value(&short).is_err());
+        let huge = msg(MessageType::FillValue, &[3, 0x29, 0xff, 0xff, 0xff, 0x7f]);
+        assert!(matches!(
+            parse_fill_value(&huge),
+            Err(FormatError::Overflow(_))
+        ));
+    }
+
+    #[test]
+    fn fill_cell_clips_edge_chunks_in_2d() {
+        // 3x5 dataset, 2x2 chunks; fill grid cell (1, 2): rows 2..3, cols 4..5.
+        let mut out = vec![0u8; 15];
+        fill_cell(&mut out, &[1, 2], &[2, 2], &[3, 5], &[5, 1], 1, &[9]);
+        let mut expected = vec![0u8; 15];
+        expected[2 * 5 + 4] = 9;
+        assert_eq!(out, expected);
+
+        // Interior cell (0, 1): rows 0..2, cols 2..4.
+        let mut out = vec![0u8; 15];
+        fill_cell(&mut out, &[0, 1], &[2, 2], &[3, 5], &[5, 1], 1, &[7]);
+        let filled: Vec<usize> = out
+            .iter()
+            .enumerate()
+            .filter(|(_, b)| **b == 7)
+            .map(|(i, _)| i)
+            .collect();
+        assert_eq!(filled, [2, 3, 7, 8]);
+    }
+}
@@ -845,9 +845,33 @@ fn shuffle_decompress(data: &[u8], element_size: usize) -> Result<Vec<u8>, Forma
    let num_elements = data.len() / element_size;
    let mut result = vec![0u8; data.len()];

-    for i in 0..num_elements {
-        for j in 0..element_size {
-            result[i * element_size + j] = data[j * num_elements + i];
+    // The shuffled stream is `element_size` byte planes of `num_elements`
+    // bytes each; un-shuffling interleaves them. This is on the read path of
+    // every compressed dataset (shuffle is applied automatically before
+    // compression). The naive `result[i * es + j] = data[j * n + i]` form does
+    // a multiply and two bounds checks per byte and defeats vectorisation;
+    // fixed-width plane arrays sliced to a common length let the compiler
+    // hoist the checks and emit interleaves for the common 4- and 8-byte
+    // element sizes.
+    fn interleave<const W: usize>(data: &[u8], n: usize, out: &mut [u8]) {
+        let planes: [&[u8]; W] = core::array::from_fn(|j| &data[j * n..(j + 1) * n]);
+        for (i, element) in out.as_chunks_mut::<W>().0.iter_mut().enumerate() {
+            for (byte, plane) in element.iter_mut().zip(&planes) {
+                *byte = plane[i];
+            }
+        }
+    }
+    match element_size {
+        2 => interleave::<2>(data, num_elements, &mut result),
+        4 => interleave::<4>(data, num_elements, &mut result),
+        8 => interleave::<8>(data, num_elements, &mut result),
+        16 => interleave::<16>(data, num_elements, &mut result),
+        _ => {
+            for (i, element) in result.chunks_exact_mut(element_size).enumerate() {
+                for (j, byte) in element.iter_mut().enumerate() {
+                    *byte = data[j * num_elements + i];
+                }
+            }
        }
    }

@@ -1045,24 +1069,30 @@ fn pcodec_compress(data: &[u8], element_size: usize) -> Result<Vec<u8>, FormatEr
    match element_size {
        4 => {
            let nums: Vec<f32> = data
-                .chunks_exact(4)
-                .map(|b| f32::from_le_bytes(b.try_into().unwrap()))
+                .as_chunks::<4>()
+                .0
+                .iter()
+                .map(|b| f32::from_le_bytes(*b))
                .collect();
            simple_compress(&nums, &config)
                .map_err(|e| FormatError::CompressionError(format!("pco: {e}")))
        }
        8 => {
            let nums: Vec<f64> = data
-                .chunks_exact(8)
-                .map(|b| f64::from_le_bytes(b.try_into().unwrap()))
+                .as_chunks::<8>()
+                .0
+                .iter()
+                .map(|b| f64::from_le_bytes(*b))
                .collect();
            simple_compress(&nums, &config)
                .map_err(|e| FormatError::CompressionError(format!("pco: {e}")))
        }
        _ => {
            let nums: Vec<u32> = data
-                .chunks_exact(4)
-                .map(|b| u32::from_le_bytes(b.try_into().unwrap()))
+                .as_chunks::<4>()
+                .0
+                .iter()
+                .map(|b| u32::from_le_bytes(*b))
                .collect();
            simple_compress(&nums, &config)
                .map_err(|e| FormatError::CompressionError(format!("pco: {e}")))
@@ -1092,11 +1122,7 @@ fn pcodec_decompress(
    } else {
        MAX_DECOMPRESS_SIZE
    };
-    let n = if element_size != 0 {
-        limit_bytes / element_size
-    } else {
-        0
-    };
+    let n = limit_bytes.checked_div(element_size).unwrap_or(0);
    match element_size {
        4 => {
            let mut buf = vec![0f32; n];
@@ -1543,8 +1569,10 @@ mod tests {

    fn as_f32(bytes: &[u8]) -> Vec<f32> {
        bytes
-            .chunks_exact(4)
-            .map(|c| f32::from_le_bytes(c.try_into().unwrap()))
+            .as_chunks::<4>()
+            .0
+            .iter()
+            .map(|c| f32::from_le_bytes(*c))
            .collect()
    }

@@ -1578,8 +1606,10 @@ mod tests {

    fn as_f64(bytes: &[u8]) -> Vec<f64> {
        bytes
-            .chunks_exact(8)
-            .map(|c| f64::from_le_bytes(c.try_into().unwrap()))
+            .as_chunks::<8>()
+            .0
+            .iter()
+            .map(|c| f64::from_le_bytes(*c))
            .collect()
    }

@@ -1842,4 +1872,20 @@ mod tests {
        };
        assert!(decompress_chunk(&data, &pipeline, 16, 1).is_err());
    }
+    #[test]
+    fn unshuffle_inverts_shuffle_for_every_element_size() {
+        for element_size in [1usize, 2, 3, 4, 5, 8, 12, 16, 24] {
+            for elements in [0usize, 1, 2, 7, 64, 1000] {
+                let original: Vec<u8> = (0..element_size * elements)
+                    .map(|i| (i * 31 + 7) as u8)
+                    .collect();
+                let shuffled = shuffle_compress(&original, element_size).unwrap();
+                assert_eq!(
+                    shuffle_decompress(&shuffled, element_size).unwrap(),
+                    original,
+                    "element_size {element_size}, {elements} elements"
+                );
+            }
+        }
+    }
 }
@@ -47,6 +47,19 @@ fn read_length(data: &[u8], pos: usize, size: u8) -> Result<u64, FormatError> {
    read_offset(data, pos, size)
 }

+fn ensure_len(data: &[u8], offset: usize, needed: usize) -> Result<(), FormatError> {
+    if offset
+        .checked_add(needed)
+        .is_none_or(|end| end > data.len())
+    {
+        return Err(FormatError::UnexpectedEof {
+            expected: offset.saturating_add(needed),
+            available: data.len(),
+        });
+    }
+    Ok(())
+}
+
 fn is_undefined(data: &[u8], pos: usize, size: u8) -> bool {
    let s = size as usize;
    if pos + s > data.len() {
@@ -66,12 +79,7 @@ impl FixedArrayHeader {
        // FAHD signature(4) + version(1) + client_id(1) + element_size(1) +
        // max_nelmts_bits(1) + num_elements(length_size) + data_block_addr(offset_size) + checksum(4)
        let min_size = 4 + 1 + 1 + 1 + 1 + length_size as usize + offset_size as usize + 4;
-        if offset + min_size > file_data.len() {
-            return Err(FormatError::UnexpectedEof {
-                expected: offset + min_size,
-                available: file_data.len(),
-            });
-        }
+        ensure_len(file_data, offset, min_size)?;

        let d = &file_data[offset..];
        if &d[0..4] != b"FAHD" {
@@ -126,12 +134,7 @@ pub fn read_fixed_array_chunks(

    // Parse data block header: FADB(4) + version(1) + client_id(1) + header_address(offset_size)
    let db_header_size = 4 + 1 + 1 + offset_size as usize;
-    if db_offset + db_header_size > file_data.len() {
-        return Err(FormatError::UnexpectedEof {
-            expected: db_offset + db_header_size,
-            available: file_data.len(),
-        });
-    }
+    ensure_len(file_data, db_offset, db_header_size)?;

    let d = &file_data[db_offset..];
    if &d[0..4] != b"FADB" {
@@ -489,6 +492,29 @@ mod tests {
        assert!(r.is_err());
    }

+    /// A near-`usize::MAX` offset must error cleanly, not overflow/panic.
+    #[test]
+    fn parse_rejects_offset_overflow() {
+        let buf = vec![0u8; 64];
+        let result = FixedArrayHeader::parse(&buf, usize::MAX - 4, 8, 8);
+        assert!(result.is_err());
+    }
+
+    /// A near-`usize::MAX` data block address must error cleanly, not overflow/panic.
+    #[test]
+    fn read_rejects_data_block_offset_overflow() {
+        let header = FixedArrayHeader {
+            client_id: 0,
+            element_size: 8,
+            max_nelmts_bits: 10,
+            num_elements: 1,
+            data_block_address: (usize::MAX - 4) as u64,
+        };
+        let buf = vec![0u8; 64];
+        let r = read_fixed_array_chunks(&buf, &header, &[100], &[20], 8, 8, 8);
+        assert!(r.is_err());
+    }
+
    #[test]
    fn parse_fixed_array_header_invalid_version() {
        let mut buf = vec![0u8; 256];
@@ -184,9 +184,7 @@ mod tests {
            buf.extend_from_slice(data);
            // Pad to 8 bytes
            let padded = pad8(data.len());
-            for _ in data.len()..padded {
-                buf.push(0);
-            }
+            buf.resize(buf.len() + (padded - data.len()), 0);
        }

        // Free space marker
@@ -60,6 +60,54 @@ pub fn resolve_v1_group_entries(
    Ok(entries)
 }

+/// Symbol table cache type for a soft link: the scratch pad's first four bytes
+/// are the local-heap offset of the link's target path, and the entry's object
+/// header address is undefined.
+const CACHE_TYPE_SOFT_LINK: u32 = 2;
+
+/// The target path of the soft link called `name` in a v1 group, if any.
+pub fn find_v1_soft_link(
+    file_data: &[u8],
+    sym_table_msg: &SymbolTableMessage,
+    name: &str,
+    offset_size: u8,
+    length_size: u8,
+) -> Result<Option<String>, FormatError> {
+    let heap = LocalHeap::parse(
+        file_data,
+        sym_table_msg.local_heap_address as usize,
+        offset_size,
+        length_size,
+    )?;
+    let snod_addrs = collect_symbol_table_nodes(
+        file_data,
+        sym_table_msg.btree_address,
+        offset_size,
+        length_size,
+    )?;
+    for snod_addr in snod_addrs {
+        let snod = SymbolTableNode::parse(file_data, snod_addr as usize, offset_size)?;
+        for entry in &snod.entries {
+            if entry.cache_type != CACHE_TYPE_SOFT_LINK {
+                continue;
+            }
+            if heap.read_string(file_data, entry.link_name_offset)? != name {
+                continue;
+            }
+            let value_offset = u32::from_le_bytes([
+                entry.scratch_pad[0],
+                entry.scratch_pad[1],
+                entry.scratch_pad[2],
+                entry.scratch_pad[3],
+            ]);
+            return heap
+                .read_string(file_data, u64::from(value_offset))
+                .map(Some);
+        }
+    }
+    Ok(None)
+}
+
 /// Extract the SymbolTableMessage from an object header's messages.
 fn find_symbol_table_message(
    obj_header: &ObjectHeader,
@@ -63,14 +63,15 @@ fn resolve_compact_entries(
    Ok(entries)
 }

-/// Resolve entries from dense storage (fractal heap + B-tree v2).
-fn resolve_dense_entries(
+/// Visit every link in dense storage (fractal heap + B-tree v2 name index).
+fn for_each_dense_link(
    file_data: &[u8],
    link_info: &LinkInfoMessage,
    fh_addr: u64,
    offset_size: u8,
    length_size: u8,
-) -> Result<Vec<GroupEntry>, FormatError> {
+    mut visit: impl FnMut(LinkMessage),
+) -> Result<(), FormatError> {
    // Parse fractal heap
    let fh = FractalHeapHeader::parse(file_data, fh_addr as usize, offset_size, length_size)?;

@@ -81,7 +82,6 @@ fn resolve_dense_entries(
    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 mut entries = Vec::new();
    for record in &records {
        // For type 5 (name index): hash(4) + heap_id(heap_id_length)
        // For type 6 (creation order): creation_order(8) + heap_id(heap_id_length)
@@ -98,22 +98,94 @@ fn resolve_dense_entries(

        // Read managed object from fractal heap
        let link_data = fh.read_managed_object(file_data, id_bytes, offset_size)?;
+        visit(LinkMessage::parse(&link_data, offset_size)?);
+    }
+    Ok(())
+}

-        // Parse as Link message
-        let link = LinkMessage::parse(&link_data, offset_size)?;
-        if let LinkTarget::Hard {
-            object_header_address,
-        } = link.link_target
-        {
-            entries.push(GroupEntry {
-                name: link.name,
+/// Resolve entries from dense storage (fractal heap + B-tree v2).
+fn resolve_dense_entries(
+    file_data: &[u8],
+    link_info: &LinkInfoMessage,
+    fh_addr: u64,
+    offset_size: u8,
+    length_size: u8,
+) -> Result<Vec<GroupEntry>, FormatError> {
+    let mut entries = Vec::new();
+    for_each_dense_link(
+        file_data,
+        link_info,
+        fh_addr,
+        offset_size,
+        length_size,
+        |link| {
+            if let LinkTarget::Hard {
                object_header_address,
-                cache_type: 0,
-            });
+            } = link.link_target
+            {
+                entries.push(GroupEntry {
+                    name: link.name,
+                    object_header_address,
+                    cache_type: 0,
+                });
+            }
+        },
+    )?;
+    Ok(entries)
+}
+
+/// The soft or external link called `name` in this group, if there is one.
+/// Hard links are what `resolve_group_entries` returns; this is consulted only
+/// when a path component isn't among them.
+fn find_symbolic_link(
+    file_data: &[u8],
+    object_header: &ObjectHeader,
+    name: &str,
+    offset_size: u8,
+    length_size: u8,
+) -> Result<Option<LinkTarget>, FormatError> {
+    if is_v1_group(object_header) {
+        let Some(sym_msg) = object_header
+            .messages
+            .iter()
+            .find(|m| m.msg_type == MessageType::SymbolTable)
+        else {
+            return Ok(None);
+        };
+        let stm = SymbolTableMessage::parse(&sym_msg.data, offset_size)?;
+        return group_v1::find_v1_soft_link(file_data, &stm, name, offset_size, length_size)
+            .map(|target| target.map(|target_path| LinkTarget::Soft { target_path }));
+    }
+    if !is_v2_group(object_header) {
+        return Ok(None);
+    }
+    let is_symbolic = |t: &LinkTarget| !matches!(t, LinkTarget::Hard { .. });
+    let link_info = find_link_info(object_header, offset_size)?;
+    let mut found = None;
+    if let Some(fh_addr) = link_info.fractal_heap_address {
+        for_each_dense_link(
+            file_data,
+            &link_info,
+            fh_addr,
+            offset_size,
+            length_size,
+            |link| {
+                if link.name == name && is_symbolic(&link.link_target) {
+                    found = Some(link.link_target);
+                }
+            },
+        )?;
+    } else {
+        for msg in &object_header.messages {
+            if msg.msg_type == MessageType::Link {
+                let link = LinkMessage::parse(&msg.data, offset_size)?;
+                if link.name == name && is_symbolic(&link.link_target) {
+                    found = Some(link.link_target);
+                }
+            }
        }
    }
-
-    Ok(entries)
+    Ok(found)
 }

 /// Find and parse the Link Info message from an object header.
@@ -158,6 +230,19 @@ pub fn resolve_path_any(
    file_data: &[u8],
    superblock: &Superblock,
    path: &str,
+) -> Result<u64, FormatError> {
+    resolve_path_following_links(file_data, superblock, path, 0)
+}
+
+/// Soft links followed while resolving one path. Guards against link cycles
+/// (`a -> b -> a`), which are legal to create.
+const MAX_SOFT_LINK_DEPTH: u8 = 16;
+
+fn resolve_path_following_links(
+    file_data: &[u8],
+    superblock: &Superblock,
+    path: &str,
+    depth: u8,
 ) -> Result<u64, FormatError> {
    let components: Vec<&str> = path.split('/').filter(|s| !s.is_empty()).collect();
    if components.is_empty() {
@@ -176,7 +261,9 @@ pub fn resolve_path_any(
    for (i, component) in components.iter().enumerate() {
        let entries = resolve_group_entries(file_data, &current_header, os, ls)?;

-        let found = entries.iter().find(|e| e.name == *component);
+        let found = entries
+            .iter()
+            .find(|e| e.name == *component && e.object_header_address != u64::MAX);
        match found {
            Some(entry) => {
                if i == components.len() - 1 {
@@ -186,7 +273,37 @@ pub fn resolve_path_any(
                current_header = ObjectHeader::parse(file_data, current_addr as usize, os, ls)?;
            }
            None => {
-                return Err(FormatError::PathNotFound(String::from(*component)));
+                return match find_symbolic_link(file_data, &current_header, component, os, ls)? {
+                    Some(LinkTarget::Soft { target_path }) => {
+                        if depth >= MAX_SOFT_LINK_DEPTH {
+                            return Err(FormatError::NestingDepthExceeded);
+                        }
+                        // A relative target is relative to the group holding
+                        // the link; then the rest of the original path.
+                        let mut full = String::new();
+                        if !target_path.starts_with('/') {
+                            for parent in &components[..i] {
+                                full.push('/');
+                                full.push_str(parent);
+                            }
+                        }
+                        full.push('/');
+                        full.push_str(&target_path);
+                        for rest in &components[i + 1..] {
+                            full.push('/');
+                            full.push_str(rest);
+                        }
+                        resolve_path_following_links(file_data, superblock, &full, depth + 1)
+                    }
+                    Some(LinkTarget::External {
+                        filename,
+                        object_path,
+                    }) => Err(FormatError::ExternalLinkUnsupported {
+                        filename,
+                        object_path,
+                    }),
+                    _ => Err(FormatError::PathNotFound(String::from(*component))),
+                };
            }
        }
    }
@@ -67,6 +67,7 @@ pub mod ea_writer;
 pub mod error;
 pub mod extensible_array;
 pub mod file_writer;
+pub mod fill_value;
 pub mod filter_pipeline;
 pub mod filters;
 mod filters_szip;
@@ -88,6 +89,7 @@ pub mod object_header;
 pub mod object_header_writer;
 #[cfg(feature = "parallel")]
 pub mod parallel_read;
+pub mod partial_read;
 pub mod profiling;
 pub mod property_list;
 pub mod selection;
@@ -413,11 +413,8 @@ mod tests {
    #[test]
    fn soft_link() {
        let target = "/group1/dataset";
-        let mut data = Vec::new();
-        data.push(1); // version
-        data.push(0x08); // flags: bit 3 = link type present, name size = 1 byte (bits 0-1 = 0)
-        data.push(1); // link type = soft
-        data.push(4); // name length = 4
+        // version, flags (bit 3 = link type present, name size = 1 byte), link type = soft, name length = 4
+        let mut data = vec![1, 0x08, 1, 4];
        data.extend_from_slice(b"link");
        data.extend_from_slice(&(target.len() as u16).to_le_bytes());
        data.extend_from_slice(target.as_bytes());
@@ -455,12 +452,8 @@ mod tests {

    #[test]
    fn invalid_link_type() {
-        let mut data = Vec::new();
-        data.push(1); // version
-        data.push(0x08); // flags: bit 3 = link type present
-        data.push(99); // invalid link type
-        data.push(1); // name length = 1
-        data.push(b'x');
+        // version, flags (bit 3 = link type present), invalid link type = 99, name length = 1, name = 'x'
+        let data = vec![1, 0x08, 99, 1, b'x'];
        let err = LinkMessage::parse(&data, 8).unwrap_err();
        assert_eq!(err, FormatError::InvalidLinkType(99));
    }
@@ -9,6 +9,9 @@ pub enum MessageType {
    Datatype,
    FillValueOld,
    FillValue,
+    /// External Data Files (0x0007): the dataset's raw data lives in other
+    /// files, listed by this message.
+    ExternalDataFiles,
    Link,
    DataLayout,
    GroupInfo,
@@ -36,6 +39,7 @@ impl MessageType {
            0x0004 => MessageType::FillValueOld,
            0x0005 => MessageType::FillValue,
            0x0006 => MessageType::Link,
+            0x0007 => MessageType::ExternalDataFiles,
            0x0008 => MessageType::DataLayout,
            0x000A => MessageType::GroupInfo,
            0x000B => MessageType::FilterPipeline,
@@ -60,6 +64,7 @@ impl MessageType {
            MessageType::Datatype => 0x0003,
            MessageType::FillValueOld => 0x0004,
            MessageType::FillValue => 0x0005,
+            MessageType::ExternalDataFiles => 0x0007,
            MessageType::Link => 0x0006,
            MessageType::DataLayout => 0x0008,
            MessageType::GroupInfo => 0x000A,
@@ -90,6 +95,7 @@ mod tests {
            (0x0003, MessageType::Datatype),
            (0x0004, MessageType::FillValueOld),
            (0x0005, MessageType::FillValue),
+            (0x0007, MessageType::ExternalDataFiles),
            (0x0006, MessageType::Link),
            (0x0008, MessageType::DataLayout),
            (0x000A, MessageType::GroupInfo),
@@ -119,8 +125,13 @@ mod tests {

    #[test]
    fn unknown_type_zero_gap() {
-        // 0x0007 is not a defined type
-        let mt = MessageType::from_u16(0x0007);
-        assert_eq!(mt, MessageType::Unknown(0x0007));
+        // 0x0009 is reserved for the library's own testing; no file uses it.
+        let mt = MessageType::from_u16(0x0009);
+        assert_eq!(mt, MessageType::Unknown(0x0009));
+        // 0x0007 used to be treated as unknown: it is External Data Files.
+        assert_eq!(
+            MessageType::from_u16(0x0007),
+            MessageType::ExternalDataFiles
+        );
    }
 }
@@ -73,9 +73,12 @@ pub fn decompress_chunks_lane_partitioned(
                let c_addr = chunk_info.address as usize;
                let size = chunk_info.chunk_size as usize;

-                if c_addr + size > file_data.len() {
+                if c_addr
+                    .checked_add(size)
+                    .is_none_or(|end| end > file_data.len())
+                {
                    return Err(FormatError::UnexpectedEof {
-                        expected: c_addr + size,
+                        expected: c_addr.saturating_add(size),
                        available: file_data.len(),
                    });
                }
@@ -144,9 +147,12 @@ pub fn decompress_chunks_parallel(
        .map(|(index, chunk_info)| {
            let c_addr = chunk_info.address as usize;
            let size = chunk_info.chunk_size as usize;
-            if c_addr + size > file_data.len() {
+            if c_addr
+                .checked_add(size)
+                .is_none_or(|end| end > file_data.len())
+            {
                return Err(FormatError::UnexpectedEof {
-                    expected: c_addr + size,
+                    expected: c_addr.saturating_add(size),
                    available: file_data.len(),
                });
            }
@@ -182,9 +188,12 @@ pub fn decompress_chunks_sequential(
    for chunk_info in chunks {
        let c_addr = chunk_info.address as usize;
        let size = chunk_info.chunk_size as usize;
-        if c_addr + size > file_data.len() {
+        if c_addr
+            .checked_add(size)
+            .is_none_or(|end| end > file_data.len())
+        {
            return Err(FormatError::UnexpectedEof {
-                expected: c_addr + size,
+                expected: c_addr.saturating_add(size),
                available: file_data.len(),
            });
        }
@@ -0,0 +1,359 @@
+//! Selection reads that cost what the selection costs, not what the dataset
+//! costs.
+//!
+//! [`crate::data_read::read_raw_data_selection`] used to decode the *entire*
+//! dataset and then pick elements out of it, so reading a 64x64 window of a
+//! large dataset took about as long as reading all of it. Here the selection's
+//! bounding box is materialised instead — only the rows of a contiguous
+//! dataset, or only the chunks, that overlap it — and the existing extractor
+//! runs over that small buffer with the selection translated to the box's
+//! origin. Extraction semantics are therefore exactly the full-read ones.
+
+#[cfg(not(feature = "std"))]
+use alloc::string as alloc_or_std;
+#[cfg(not(feature = "std"))]
+use alloc::{format, vec, vec::Vec};
+#[cfg(feature = "std")]
+use std::string as alloc_or_std;
+
+use crate::chunked_read::{alloc_output, checked_byte_len, list_chunks};
+use crate::data_layout::DataLayout;
+use crate::data_read::extract_selection_from_buffer;
+use crate::dataspace::Dataspace;
+use crate::error::FormatError;
+use crate::filter_pipeline::FilterPipeline;
+use crate::filters::decompress_chunk;
+use crate::selection::Selection;
+
+/// The smallest axis-aligned box containing every selected element, as
+/// `(start, extent)` per dimension. `None` when there is nothing to gain or
+/// the selection is not valid for `dims` (the caller's full path then reports
+/// the error exactly as before).
+fn bounding_box(selection: &Selection, dims: &[u64]) -> Option<(Vec<u64>, Vec<u64>)> {
+    match selection {
+        Selection::Hyperslab {
+            start,
+            stride,
+            count,
+            block,
+        } => {
+            let rank = dims.len();
+            if [start.len(), stride.len(), count.len(), block.len()] != [rank; 4] {
+                return None;
+            }
+            let mut extent = Vec::with_capacity(rank);
+            for d in 0..rank {
+                if count[d] == 0 || block[d] == 0 {
+                    return None;
+                }
+                // Last selected index + 1, relative to start.
+                let span = (count[d] - 1)
+                    .checked_mul(stride[d])?
+                    .checked_add(block[d])?;
+                if start[d].checked_add(span)? > dims[d] {
+                    return None;
+                }
+                extent.push(span);
+            }
+            Some((start.clone(), extent))
+        }
+        Selection::Points(points) => {
+            let rank = dims.len();
+            let first = points.first()?;
+            if first.len() != rank {
+                return None;
+            }
+            let (mut lo, mut hi) = (first.clone(), first.clone());
+            for p in points {
+                if p.len() != rank {
+                    return None;
+                }
+                for d in 0..rank {
+                    if p[d] >= dims[d] {
+                        return None;
+                    }
+                    lo[d] = lo[d].min(p[d]);
+                    hi[d] = hi[d].max(p[d]);
+                }
+            }
+            let extent = lo.iter().zip(&hi).map(|(l, h)| h - l + 1).collect();
+            Some((lo, extent))
+        }
+        Selection::All | Selection::None => None,
+    }
+}
+
+/// Check that `selection` addresses only elements that exist in a dataset of
+/// shape `dims`. Without this an out-of-range selection read *something*: a
+/// hyperslab past the edge came back padded with zeros, and a point whose
+/// column was out of range wrapped into the next row.
+pub fn validate(selection: &Selection, dims: &[u64]) -> Result<(), FormatError> {
+    let rank = dims.len();
+    let bad = |msg: alloc_or_std::String| Err(FormatError::SelectionOutOfBounds(msg));
+    match selection {
+        Selection::All | Selection::None => Ok(()),
+        Selection::Hyperslab {
+            start,
+            stride,
+            count,
+            block,
+        } => {
+            if [start.len(), stride.len(), count.len(), block.len()] != [rank; 4] {
+                return bad(format!("hyperslab rank does not match dataset rank {rank}"));
+            }
+            for d in 0..rank {
+                if count[d] == 0 || block[d] == 0 {
+                    continue; // selects nothing along this dimension
+                }
+                let end = (count[d] - 1)
+                    .checked_mul(stride[d])
+                    .and_then(|v| v.checked_add(block[d]))
+                    .and_then(|v| v.checked_add(start[d]));
+                if !end.is_some_and(|end| end <= dims[d]) {
+                    return bad(format!(
+                        "dimension {d}: start {} stride {} count {} block {} exceeds extent {}",
+                        start[d], stride[d], count[d], block[d], dims[d]
+                    ));
+                }
+                if block[d] > stride[d] && count[d] > 1 {
+                    return bad(format!(
+                        "dimension {d}: block {} larger than stride {} (overlapping blocks)",
+                        block[d], stride[d]
+                    ));
+                }
+            }
+            Ok(())
+        }
+        Selection::Points(points) => {
+            for p in points {
+                if p.len() != rank {
+                    return bad(format!("point {p:?} does not match dataset rank {rank}"));
+                }
+                if let Some(d) = (0..rank).find(|&d| p[d] >= dims[d]) {
+                    return bad(format!(
+                        "point {p:?}: coordinate {} exceeds extent {} of dimension {d}",
+                        p[d], dims[d]
+                    ));
+                }
+            }
+            Ok(())
+        }
+    }
+}
+
+/// The same selection expressed relative to `origin`.
+fn translate(selection: &Selection, origin: &[u64]) -> Selection {
+    match selection {
+        Selection::Hyperslab {
+            start,
+            stride,
+            count,
+            block,
+        } => Selection::Hyperslab {
+            start: start.iter().zip(origin).map(|(s, o)| s - o).collect(),
+            stride: stride.clone(),
+            count: count.clone(),
+            block: block.clone(),
+        },
+        Selection::Points(points) => Selection::Points(
+            points
+                .iter()
+                .map(|p| p.iter().zip(origin).map(|(c, o)| c - o).collect())
+                .collect(),
+        ),
+        other => other.clone(),
+    }
+}
+
+/// Copy the part of a source region that overlaps the box into `out` (which
+/// is the box, row-major).
+///
+/// The source region starts at `src_origin` in dataset coordinates, has shape
+/// `src_shape`, and its elements are in `src` row-major. One `memcpy` per
+/// overlapping row of the last dimension.
+#[allow(clippy::too_many_arguments)]
+fn copy_overlap(
+    src: &[u8],
+    src_origin: &[u64],
+    src_shape: &[u64],
+    out: &mut [u8],
+    box_start: &[u64],
+    box_extent: &[u64],
+    elem_size: usize,
+) {
+    let rank = box_start.len();
+    // Overlap in dataset coordinates.
+    let mut lo = vec![0u64; rank];
+    let mut hi = vec![0u64; rank];
+    for d in 0..rank {
+        lo[d] = src_origin[d].max(box_start[d]);
+        hi[d] = (src_origin[d] + src_shape[d]).min(box_start[d] + box_extent[d]);
+        if lo[d] >= hi[d] {
+            return;
+        }
+    }
+    let strides = |shape: &[u64]| {
+        let mut s = vec![1u64; rank];
+        for d in (0..rank.saturating_sub(1)).rev() {
+            s[d] = s[d + 1] * shape[d + 1];
+        }
+        s
+    };
+    let (src_strides, out_strides) = (strides(src_shape), strides(box_extent));
+    let last = rank - 1;
+    let run = ((hi[last] - lo[last]) as usize) * elem_size;
+
+    let mut idx = lo.clone();
+    loop {
+        let src_at: u64 = (0..rank)
+            .map(|d| (idx[d] - src_origin[d]) * src_strides[d])
+            .sum();
+        let out_at: u64 = (0..rank)
+            .map(|d| (idx[d] - box_start[d]) * out_strides[d])
+            .sum();
+        let (s, o) = (src_at as usize * elem_size, out_at as usize * elem_size);
+        if let (Some(from), Some(to)) = (src.get(s..s + run), out.get_mut(o..o + run)) {
+            to.copy_from_slice(from);
+        }
+        // Advance over every dimension but the last.
+        let mut d = last;
+        loop {
+            if d == 0 {
+                return;
+            }
+            d -= 1;
+            idx[d] += 1;
+            if idx[d] < hi[d] {
+                break;
+            }
+            idx[d] = lo[d];
+        }
+    }
+}
+
+/// Read `selection` without materialising the whole dataset, when that is
+/// possible and worthwhile. `Ok(None)` means "use the full-read path": an
+/// `All`/`None`/invalid selection, a layout this doesn't handle (compact,
+/// virtual, storage-less), or a bounding box covering most of the dataset.
+#[allow(clippy::too_many_arguments)]
+pub fn read_selection(
+    file_data: &[u8],
+    layout: &DataLayout,
+    dataspace: &Dataspace,
+    elem_size: usize,
+    pipeline: Option<&FilterPipeline>,
+    offset_size: u8,
+    length_size: u8,
+    selection: &Selection,
+) -> Result<Option<Vec<u8>>, FormatError> {
+    let dims = &dataspace.dimensions;
+    if dims.is_empty() || elem_size == 0 {
+        return Ok(None);
+    }
+    let Some((box_start, box_extent)) = bounding_box(selection, dims) else {
+        return Ok(None);
+    };
+    let total = dataspace.checked_num_elements()?;
+    let box_elements = box_extent
+        .iter()
+        .try_fold(1u64, |acc, &e| acc.checked_mul(e))
+        .ok_or_else(|| FormatError::Overflow("selection bounding box overflows".into()))?;
+    // A box covering most of the dataset gains nothing over the full path.
+    if box_elements.saturating_mul(2) > total {
+        return Ok(None);
+    }
+    let mut boxed = alloc_output(checked_byte_len(box_elements, elem_size)?)?;
+
+    match layout {
+        DataLayout::Contiguous {
+            address: Some(address),
+            ..
+        } => {
+            let base = usize::try_from(*address)
+                .map_err(|_| FormatError::Overflow("data address exceeds usize".into()))?;
+            let data = file_data
+                .get(base..)
+                .and_then(|d| d.get(..checked_byte_len(total, elem_size).ok()?))
+                .ok_or(FormatError::UnexpectedEof {
+                    expected: base,
+                    available: file_data.len(),
+                })?;
+            let origin = vec![0u64; dims.len()];
+            copy_overlap(
+                data,
+                &origin,
+                dims,
+                &mut boxed,
+                &box_start,
+                &box_extent,
+                elem_size,
+            );
+        }
+        DataLayout::Chunked {
+            btree_address: Some(_),
+            ..
+        } => {
+            let (chunks, chunk_dims) = list_chunks(
+                file_data,
+                layout,
+                dataspace,
+                elem_size,
+                offset_size,
+                length_size,
+            )?;
+            let rank = dims.len();
+            let chunk_shape: Vec<u64> = chunk_dims.iter().map(|&d| d as u64).collect();
+            let chunk_bytes = crate::chunked_read::checked_chunk_byte_len(&chunk_dims, elem_size)?;
+            for chunk in &chunks {
+                if chunk.offsets.len() < rank || chunk.address == u64::MAX {
+                    continue;
+                }
+                let origin = &chunk.offsets[..rank];
+                let overlaps = (0..rank).all(|d| {
+                    origin[d] < box_start[d] + box_extent[d]
+                        && origin[d].saturating_add(chunk_shape[d]) > box_start[d]
+                });
+                if !overlaps {
+                    continue;
+                }
+                let at = usize::try_from(chunk.address)
+                    .map_err(|_| FormatError::Overflow("chunk address exceeds usize".into()))?;
+                let raw = at
+                    .checked_add(chunk.chunk_size as usize)
+                    .and_then(|end| file_data.get(at..end))
+                    .ok_or(FormatError::UnexpectedEof {
+                        expected: at.saturating_add(chunk.chunk_size as usize),
+                        available: file_data.len(),
+                    })?;
+                // Mirrors the full-read path: a non-zero filter mask means the
+                // chunk was stored unfiltered.
+                let decoded;
+                let data: &[u8] = match pipeline {
+                    Some(pl) if chunk.filter_mask == 0 => {
+                        decoded = decompress_chunk(raw, pl, chunk_bytes, elem_size as u32)?;
+                        &decoded
+                    }
+                    _ => raw,
+                };
+                copy_overlap(
+                    data,
+                    origin,
+                    &chunk_shape,
+                    &mut boxed,
+                    &box_start,
+                    &box_extent,
+                    elem_size,
+                );
+            }
+        }
+        _ => return Ok(None),
+    }
+
+    extract_selection_from_buffer(
+        &boxed,
+        &box_extent,
+        elem_size,
+        &translate(selection, &box_start),
+    )
+    .map(Some)
+}
@@ -509,7 +509,7 @@ mod tests {

    #[test]
    fn selection_slice_1d() {
-        let sel = Selection::slice(&[5..15]);
+        let sel = Selection::slice(std::slice::from_ref(&(5..15)));
        assert_eq!(sel.num_elements(&[100]), 10);
        assert_eq!(sel.output_shape(&[100]), vec![10]);
    }
@@ -16,8 +16,12 @@
 //! - SMLI list structure: simple list of shared message entries
 //! - B-tree v2 type 7: indexed shared message entries

+#[cfg(not(feature = "std"))]
+use alloc::borrow::Cow;
 #[cfg(not(feature = "std"))]
 use alloc::vec::Vec;
+#[cfg(feature = "std")]
+use std::borrow::Cow;

 use crate::btree_v2::{BTreeV2Header, collect_btree_v2_records};
 use crate::error::FormatError;
@@ -28,6 +32,14 @@ use crate::object_header::ObjectHeader;
 /// Fractal heap ID length for SOHM entries (fixed at 8 bytes).
 const FHEAP_ID_LEN: usize = 8;

+/// Shared-message `type` values (version 3 encoding).
+/// The message is in the file's shared-message (SOHM) fractal heap.
+const SHARE_TYPE_SOHM: u8 = 1;
+/// The message is in another object's header (a committed/named datatype).
+const SHARE_TYPE_COMMITTED: u8 = 2;
+/// The message is stored here but is sharable.
+const SHARE_TYPE_HERE: u8 = 3;
+
 /// A resolved shared message reference.
 #[derive(Debug, Clone)]
 pub struct SharedMessageRef {
@@ -35,9 +47,10 @@ pub struct SharedMessageRef {
    pub ref_type: u8,
    /// Version of the shared message encoding.
    pub version: u8,
-    /// Address of the object header containing the shared message (type 1, 3).
+    /// Address of the object header holding the message (committed). Set for
+    /// every v1/v2 reference and for v3 types 2 and 3.
    pub object_header_address: Option<u64>,
-    /// Fractal heap ID for type 2 (SOHM) references.
+    /// Fractal heap ID for a v3 SOHM (type 1) reference.
    pub heap_id: Option<[u8; FHEAP_ID_LEN]>,
 }

@@ -146,48 +159,39 @@ pub fn parse_shared_ref(data: &[u8], offset_size: u8) -> Result<SharedMessageRef
    let version = data[0];
    let ref_type = data[1];

+    // Layouts (HDF5 spec IV.A.2 "Shared Message", and libhdf5's decoder):
+    //   v1: version, type, reserved(6), address     — always "committed"
+    //   v2: version, type, address                  — always "committed"
+    //   v3: version, type, then a fractal-heap ID if type == SOHM, otherwise
+    //       an address
+    // Verified against h5py/HDF5 2.0 output, which writes `02 02 <address>`
+    // for a dataset using a committed datatype under both default and
+    // `latest` libver bounds.
+    let address_at = |pos: usize| -> Result<SharedMessageRef, FormatError> {
+        ensure_len(data, pos, offset_size as usize)?;
+        Ok(SharedMessageRef {
+            ref_type,
+            version,
+            object_header_address: Some(read_offset(data, pos, offset_size)?),
+            heap_id: None,
+        })
+    };
    match version {
-        1 | 2 => {
-            // v1/v2: reserved(6) + address(offset_size)
-            let pos = 2 + 6; // skip reserved bytes
-            ensure_len(data, pos, offset_size as usize)?;
-            let addr = read_offset(data, pos, offset_size)?;
+        1 => address_at(2 + 6),
+        2 => address_at(2),
+        3 if ref_type == SHARE_TYPE_SOHM => {
+            ensure_len(data, 2, FHEAP_ID_LEN)?;
+            let mut id = [0u8; FHEAP_ID_LEN];
+            id.copy_from_slice(&data[2..2 + FHEAP_ID_LEN]);
            Ok(SharedMessageRef {
                ref_type,
                version,
-                object_header_address: Some(addr),
-                heap_id: None,
+                object_header_address: None,
+                heap_id: Some(id),
            })
        }
-        3 => {
-            match ref_type {
-                1 | 3 => {
-                    // type 1/3: message in another object header
-                    // v3 layout: version(1) + type(1) + address(offset_size)
-                    ensure_len(data, 2, offset_size as usize)?;
-                    let addr = read_offset(data, 2, offset_size)?;
-                    Ok(SharedMessageRef {
-                        ref_type,
-                        version,
-                        object_header_address: Some(addr),
-                        heap_id: None,
-                    })
-                }
-                2 => {
-                    // type 2: SOHM table (fractal heap ID)
-                    ensure_len(data, 2, FHEAP_ID_LEN)?;
-                    let mut id = [0u8; FHEAP_ID_LEN];
-                    id.copy_from_slice(&data[2..2 + FHEAP_ID_LEN]);
-                    Ok(SharedMessageRef {
-                        ref_type,
-                        version,
-                        object_header_address: None,
-                        heap_id: Some(id),
-                    })
-                }
-                _ => Err(FormatError::InvalidSharedMessageVersion(ref_type)),
-            }
-        }
+        3 if ref_type == SHARE_TYPE_COMMITTED || ref_type == SHARE_TYPE_HERE => address_at(2),
+        3 => Err(FormatError::InvalidSharedMessageVersion(ref_type)),
        _ => Err(FormatError::InvalidSharedMessageVersion(version)),
    }
 }
@@ -422,6 +426,35 @@ pub fn resolve_sohm_message(
    fh_header.read_managed_object(file_data, heap_id, offset_size)
 }

+/// The payload of an object-header message, following the indirection if the
+/// message is *shared* (header flag bit 1).
+///
+/// A shared message's bytes are not the message itself but a reference to
+/// where it lives — e.g. a dataset created with a committed (named) datatype
+/// stores only a pointer to that datatype's object header. Every reader of a
+/// message that may be shared (datatype, dataspace, fill value, filter
+/// pipeline, attribute) must go through this; parsing the reference bytes as
+/// the message yields garbage rather than an error.
+pub fn message_data<'a>(
+    file_data: &[u8],
+    msg: &'a crate::object_header::HeaderMessage,
+    offset_size: u8,
+    length_size: u8,
+) -> Result<Cow<'a, [u8]>, FormatError> {
+    if !is_shared(msg.flags) {
+        return Ok(Cow::Borrowed(&msg.data));
+    }
+    let shared_ref = parse_shared_ref(&msg.data, offset_size)?;
+    resolve_shared_message(
+        file_data,
+        &shared_ref,
+        msg.msg_type,
+        offset_size,
+        length_size,
+    )
+    .map(Cow::Owned)
+}
+
 /// Resolve a shared message to its actual message data.
 ///
 /// For type 1/3 (shared in another object header), reads the target object header
@@ -453,14 +486,14 @@ pub fn resolve_shared_message_with_sohm(
    length_size: u8,
    sohm_table: Option<&SohmTable>,
 ) -> Result<Vec<u8>, FormatError> {
-    match shared_ref.ref_type {
-        1 | 3 => {
-            let addr = shared_ref
-                .object_header_address
-                .ok_or(FormatError::UnexpectedEof {
-                    expected: 1,
-                    available: 0,
-                })?;
+    // Dispatch on what the reference carries rather than on `ref_type`: v1/v2
+    // references are always an object-header address whatever their type
+    // byte says.
+    match (
+        shared_ref.object_header_address,
+        shared_ref.heap_id.as_ref(),
+    ) {
+        (Some(addr), _) => {
            let target_header =
                ObjectHeader::parse(file_data, addr as usize, offset_size, length_size)?;
            for msg in &target_header.messages {
@@ -487,11 +520,7 @@ pub fn resolve_shared_message_with_sohm(
                available: 0,
            })
        }
-        2 => {
-            let heap_id = shared_ref
-                .heap_id
-                .as_ref()
-                .ok_or(FormatError::InvalidSharedMessageVersion(2))?;
+        (None, Some(heap_id)) => {
            let table = sohm_table.ok_or(FormatError::InvalidSharedMessageVersion(2))?;
            resolve_sohm_message(
                file_data,
@@ -502,7 +531,7 @@ pub fn resolve_shared_message_with_sohm(
                length_size,
            )
        }
-        _ => Err(FormatError::InvalidSharedMessageVersion(
+        (None, None) => Err(FormatError::InvalidSharedMessageVersion(
            shared_ref.ref_type,
        )),
    }
@@ -522,15 +551,15 @@ mod tests {
    }

    #[test]
-    fn parse_v3_type1_ref() {
+    fn parse_v3_committed_ref() {
        let mut data = Vec::new();
        data.push(3); // version
-        data.push(1); // type 1 = shared in another OH
+        data.push(SHARE_TYPE_COMMITTED); // message lives in another object header
        data.extend_from_slice(&0x1234u64.to_le_bytes()); // address

        let shared = parse_shared_ref(&data, 8).unwrap();
        assert_eq!(shared.version, 3);
-        assert_eq!(shared.ref_type, 1);
+        assert_eq!(shared.ref_type, SHARE_TYPE_COMMITTED);
        assert_eq!(shared.object_header_address, Some(0x1234));
        assert!(shared.heap_id.is_none());
    }
@@ -539,7 +568,7 @@ mod tests {
    fn parse_v3_type3_ref() {
        let mut data = Vec::new();
        data.push(3); // version
-        data.push(3); // type 3 = shared in another OH (v3 encoding)
+        data.push(SHARE_TYPE_HERE); // stored here but sharable: an address
        data.extend_from_slice(&0xABCDu64.to_le_bytes());

        let shared = parse_shared_ref(&data, 8).unwrap();
@@ -563,10 +592,10 @@ mod tests {

    #[test]
    fn parse_v2_ref() {
+        // v2 dropped v1's six reserved bytes: the address follows the type.
        let mut data = Vec::new();
        data.push(2); // version
-        data.push(0); // type
-        data.extend_from_slice(&[0u8; 6]); // reserved
+        data.push(SHARE_TYPE_COMMITTED);
        data.extend_from_slice(&0x9000u32.to_le_bytes());

        let shared = parse_shared_ref(&data, 4).unwrap();
@@ -575,15 +604,26 @@ mod tests {
    }

    #[test]
-    fn parse_v3_type2_sohm() {
+    fn parse_v2_ref_from_hdf5_2_0() {
+        // Datatype message of a dataset created with a committed datatype,
+        // as written by h5py 3.16 / HDF5 2.0 (libver='latest'): header flags
+        // 0x03 (shared), payload `02 02 <8-byte object header address>`.
+        let data = [0x02, 0x02, 0xb3, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00];
+        let shared = parse_shared_ref(&data, 8).unwrap();
+        assert_eq!(shared.object_header_address, Some(0xb3));
+        assert!(shared.heap_id.is_none());
+    }
+
+    #[test]
+    fn parse_v3_sohm_ref() {
        let mut data = Vec::new();
        data.push(3); // version
-        data.push(2); // type 2 = SOHM heap
+        data.push(SHARE_TYPE_SOHM); // message lives in the SOHM fractal heap
        data.extend_from_slice(&[0xAA, 0xBB, 0xCC, 0xDD, 0x11, 0x22, 0x33, 0x44]);

        let shared = parse_shared_ref(&data, 8).unwrap();
        assert_eq!(shared.version, 3);
-        assert_eq!(shared.ref_type, 2);
+        assert_eq!(shared.ref_type, SHARE_TYPE_SOHM);
        assert_eq!(shared.object_header_address, None);
        assert_eq!(
            shared.heap_id,
@@ -592,10 +632,10 @@ mod tests {
    }

    #[test]
-    fn parse_v3_type2_too_short() {
+    fn parse_v3_sohm_too_short() {
        let mut data = Vec::new();
        data.push(3); // version
-        data.push(2); // type 2 = SOHM heap
+        data.push(SHARE_TYPE_SOHM);
        data.extend_from_slice(&[0xAA, 0xBB]); // only 2 bytes, need 8

        let err = parse_shared_ref(&data, 8).unwrap_err();
@@ -620,7 +660,7 @@ mod tests {
    fn parse_four_byte_offsets() {
        let mut data = Vec::new();
        data.push(3); // version
-        data.push(1); // type 1
+        data.push(SHARE_TYPE_COMMITTED);
        data.extend_from_slice(&0x1000u32.to_le_bytes());

        let shared = parse_shared_ref(&data, 4).unwrap();
@@ -80,9 +80,12 @@ impl SymbolTableNode {
        offset_size: u8,
    ) -> Result<SymbolTableNode, FormatError> {
        // signature(4) + version(1) + reserved(1) + number_of_symbols(2) = 8
-        if offset + 8 > file_data.len() {
+        if offset
+            .checked_add(8)
+            .is_none_or(|end| end > file_data.len())
+        {
            return Err(FormatError::UnexpectedEof {
-                expected: offset + 8,
+                expected: offset.saturating_add(8),
                available: file_data.len(),
            });
        }
@@ -103,7 +106,12 @@ impl SymbolTableNode {
        // Each entry: link_name_offset(os) + obj_hdr_addr(os) + cache_type(4) + reserved(4) + scratch(16)
        let entry_size = os + os + 4 + 4 + 16;
        let entries_start = offset + 8;
-        let needed = entries_start + num_symbols * entry_size;
+        let needed = entries_start.checked_add(num_symbols * entry_size).ok_or(
+            FormatError::UnexpectedEof {
+                expected: usize::MAX,
+                available: file_data.len(),
+            },
+        )?;
        if needed > file_data.len() {
            return Err(FormatError::UnexpectedEof {
                expected: needed,
@@ -228,4 +236,24 @@ mod tests {
        let err = SymbolTableNode::parse(&data, 0, 8).unwrap_err();
        assert_eq!(err, FormatError::InvalidSymbolTableNodeVersion(2));
    }
+
+    /// A near-`usize::MAX` SNOD offset must error cleanly, not overflow/panic.
+    #[test]
+    fn parse_snod_rejects_offset_overflow() {
+        let data = build_snod(&[], 8);
+        let result = SymbolTableNode::parse(&data, usize::MAX - 4, 8);
+        assert!(result.is_err());
+    }
+
+    /// A huge symbol count combined with a large entries_start must not
+    /// overflow the `needed` size computation.
+    #[test]
+    fn parse_snod_rejects_entries_size_overflow() {
+        let mut data = build_snod(&[], 8);
+        // num_symbols at offset 6..8 — set to max to blow up entries_start + num_symbols*entry_size
+        data[6] = 0xFF;
+        data[7] = 0xFF;
+        let result = SymbolTableNode::parse(&data, usize::MAX / 2, 8);
+        assert!(result.is_err());
+    }
 }
@@ -279,6 +279,43 @@ pub(crate) fn build_attr_message(name: &str, value: &AttrValue) -> AttributeMess
            dataspace: scalar_ds(),
            raw_data: v.to_le_bytes().to_vec(),
        },
+        AttrValue::U64Array(arr) => {
+            let mut raw = Vec::with_capacity(arr.len() * 8);
+            for v in arr {
+                raw.extend_from_slice(&v.to_le_bytes());
+            }
+            AttributeMessage {
+                name: name.to_string(),
+                datatype: Datatype::FixedPoint {
+                    size: 8,
+                    byte_order: DatatypeByteOrder::LittleEndian,
+                    signed: false,
+                    bit_offset: 0,
+                    bit_precision: 64,
+                },
+                dataspace: simple_1d(arr.len() as u64),
+                raw_data: raw,
+            }
+        }
+        AttrValue::Raw {
+            datatype,
+            shape,
+            data,
+        } => AttributeMessage {
+            name: name.to_string(),
+            datatype: datatype.clone(),
+            dataspace: if shape.is_empty() {
+                scalar_ds()
+            } else {
+                Dataspace {
+                    space_type: DataspaceType::Simple,
+                    rank: shape.len() as u8,
+                    dimensions: shape.clone(),
+                    max_dimensions: None,
+                }
+            },
+            raw_data: data.clone(),
+        },
        AttrValue::String(s) => {
            let bytes = s.as_bytes();
            AttributeMessage {
@@ -334,7 +371,7 @@ pub(crate) fn simple_1d(n: u64) -> Dataspace {

 // ---- Attribute values ----

-/// Convenient attribute values for the write API.
+/// Attribute values, for both the write API and what reading returns.
 #[derive(Debug, Clone)]
 pub enum AttrValue {
    F64(f64),
@@ -342,8 +379,21 @@ pub enum AttrValue {
    I64(i64),
    I64Array(Vec<i64>),
    U64(u64),
+    /// Unsigned integers, kept unsigned so values above `i64::MAX` survive.
+    U64Array(Vec<u64>),
    String(String),
    StringArray(Vec<String>),
+    /// An attribute whose datatype has no dedicated variant above (compound,
+    /// general enum, complex, reference, opaque, array, ...), carried verbatim
+    /// so it is never silently lost: the datatype, the dataspace dimensions
+    /// (empty for a scalar) and the element bytes exactly as stored. Decode
+    /// `data` with `clawhdf5_format::data_read` (e.g. `read_compound_fields`)
+    /// against `datatype`. Writing a `Raw` value stores it back unchanged.
+    Raw {
+        datatype: Datatype,
+        shape: Vec<u64>,
+        data: Vec<u8>,
+    },
 }

 // ---- Dataset builder ----
@@ -0,0 +1,44 @@
+"""Generate std_ref_hdf5_2_0.h5: a dataset of H5T_STD_REF (the reference
+datatype introduced in HDF5 1.12, datatype message version 4) holding two
+object references — to /target (a dataset) and /grp (a group).
+
+h5py has no API for this type, so the file is written by calling the libhdf5
+bundled in the h5py wheel directly through ctypes. Written with h5py 3.16.0 /
+HDF5 2.0.0. Re-run only if the fixture ever needs regenerating:
+
+    python gen_std_ref.py std_ref_hdf5_2_0.h5
+"""
+import ctypes
+import glob
+import os
+import sys
+
+import h5py
+import numpy as np
+
+libdir = os.path.join(os.path.dirname(os.path.dirname(h5py.__file__)), "h5py.libs")
+libs = [p for p in glob.glob(os.path.join(libdir, "libhdf5*.so*")) if "_hl" not in os.path.basename(p)]
+lib = ctypes.CDLL(libs[0])
+lib.H5open()
+hid = ctypes.c_int64
+std_ref = hid.in_dll(lib, "H5T_STD_REF_g").value
+
+lib.H5Screate_simple.restype = hid
+lib.H5Screate_simple.argtypes = [ctypes.c_int, ctypes.POINTER(ctypes.c_uint64), ctypes.POINTER(ctypes.c_uint64)]
+lib.H5Dcreate2.restype = hid
+lib.H5Dcreate2.argtypes = [hid, ctypes.c_char_p, hid, hid, hid, hid, hid]
+lib.H5Rcreate_object.argtypes = [hid, ctypes.c_char_p, hid, ctypes.c_void_p]
+lib.H5Dwrite.argtypes = [hid, hid, hid, hid, hid, ctypes.c_void_p]
+lib.H5Dclose.argtypes = [hid]
+
+with h5py.File(sys.argv[1], "w", libver="latest") as f:
+    f.create_dataset("target", data=np.arange(5, dtype="<i4"))
+    f.create_group("grp")
+    fid = f.id.id
+    sid = lib.H5Screate_simple(1, (ctypes.c_uint64 * 1)(2), None)
+    did = lib.H5Dcreate2(fid, b"refs", std_ref, sid, 0, 0, 0)
+    refs = ((ctypes.c_ubyte * 64) * 2)()  # H5R_ref_t is a 64-byte buffer
+    assert lib.H5Rcreate_object(fid, b"/target", 0, ctypes.byref(refs[0])) == 0
+    assert lib.H5Rcreate_object(fid, b"/grp", 0, ctypes.byref(refs[1])) == 0
+    assert lib.H5Dwrite(did, std_ref, 0, 0, 0, ctypes.byref(refs)) == 0
+    lib.H5Dclose(did)
@@ -343,7 +343,11 @@ fn attrs_h5_dataset_scale() {
    let scale_attr = find_attribute(&attrs, "scale").expect("scale attr not found");
    let vals = scale_attr.read_as_f64().unwrap();
    assert_eq!(vals.len(), 1);
-    assert!((vals[0] - 3.14).abs() < 1e-10);
+    // 3.14 here is the literal value baked into the binary fixture (fixtures/attrs.h5),
+    // not an arbitrary sample value, so it cannot be swapped for another constant.
+    #[allow(clippy::approx_constant)]
+    let expected = 3.14;
+    assert!((vals[0] - expected).abs() < 1e-10);
 }

 #[test]
@@ -556,8 +560,8 @@ fn chunked_deflate_read_values() {
    let (raw, datatype, _) = read_chunked_dataset(file_data, "data");
    let values = read_as_f64(&raw, &datatype).unwrap();
    assert_eq!(values.len(), 100);
-    for i in 0..100 {
-        assert_eq!(values[i], i as f64, "mismatch at index {i}");
+    for (i, &v) in values.iter().enumerate() {
+        assert_eq!(v, i as f64, "mismatch at index {i}");
    }
 }

@@ -567,8 +571,8 @@ fn chunked_shuffle_deflate_read_values() {
    let (raw, datatype, _) = read_chunked_dataset(file_data, "data");
    let values = read_as_f64(&raw, &datatype).unwrap();
    assert_eq!(values.len(), 100);
-    for i in 0..100 {
-        assert_eq!(values[i], i as f64, "mismatch at index {i}");
+    for (i, &v) in values.iter().enumerate() {
+        assert_eq!(v, i as f64, "mismatch at index {i}");
    }
 }

@@ -578,8 +582,8 @@ fn chunked_fletcher32_read_values() {
    let (raw, datatype, _) = read_chunked_dataset(file_data, "data");
    let values = read_as_f64(&raw, &datatype).unwrap();
    assert_eq!(values.len(), 100);
-    for i in 0..100 {
-        assert_eq!(values[i], i as f64, "mismatch at index {i}");
+    for (i, &v) in values.iter().enumerate() {
+        assert_eq!(v, i as f64, "mismatch at index {i}");
    }
 }

@@ -589,11 +593,10 @@ fn chunked_2d_read_values() {
    let (raw, datatype, _) = read_chunked_dataset(file_data, "matrix");
    let values = read_as_f32(&raw, &datatype).unwrap();
    assert_eq!(values.len(), 60);
-    for i in 0..60 {
+    for (i, &v) in values.iter().enumerate() {
        assert!(
-            (values[i] - i as f32).abs() < 1e-6,
-            "mismatch at index {i}: got {}",
-            values[i]
+            (v - i as f32).abs() < 1e-6,
+            "mismatch at index {i}: got {v}"
        );
    }
 }
@@ -604,8 +607,8 @@ fn chunked_large_read_values() {
    let (raw, datatype, _) = read_chunked_dataset(file_data, "big");
    let values = read_as_i32(&raw, &datatype).unwrap();
    assert_eq!(values.len(), 1000);
-    for i in 0..1000 {
-        assert_eq!(values[i], i as i32, "mismatch at index {i}");
+    for (i, &v) in values.iter().enumerate() {
+        assert_eq!(v, i as i32, "mismatch at index {i}");
    }
 }

@@ -615,8 +618,8 @@ fn chunked_nofilter_read_values() {
    let (raw, datatype, _) = read_chunked_dataset(file_data, "raw");
    let values = read_as_f64(&raw, &datatype).unwrap();
    assert_eq!(values.len(), 50);
-    for i in 0..50 {
-        assert_eq!(values[i], i as f64, "mismatch at index {i}");
+    for (i, &v) in values.iter().enumerate() {
+        assert_eq!(v, i as f64, "mismatch at index {i}");
    }
 }

@@ -646,8 +649,8 @@ fn v4_implicit_read() {
    let (raw, datatype, _) = read_chunked_dataset(file_data, "data");
    let values = read_as_f64(&raw, &datatype).unwrap();
    assert_eq!(values.len(), 100);
-    for i in 0..100 {
-        assert_eq!(values[i], i as f64, "mismatch at index {i}");
+    for (i, &v) in values.iter().enumerate() {
+        assert_eq!(v, i as f64, "mismatch at index {i}");
    }
 }

@@ -657,8 +660,8 @@ fn v4_fixed_array_read() {
    let (raw, datatype, _) = read_chunked_dataset(file_data, "data");
    let values = read_as_f64(&raw, &datatype).unwrap();
    assert_eq!(values.len(), 100);
-    for i in 0..100 {
-        assert_eq!(values[i], i as f64, "mismatch at index {i}");
+    for (i, &v) in values.iter().enumerate() {
+        assert_eq!(v, i as f64, "mismatch at index {i}");
    }
 }

@@ -871,11 +874,10 @@ fn v4_2d_fixed_array_read() {
    let (raw, datatype, _) = read_chunked_dataset(file_data, "matrix");
    let values = read_as_f32(&raw, &datatype).unwrap();
    assert_eq!(values.len(), 60);
-    for i in 0..60 {
+    for (i, &v) in values.iter().enumerate() {
        assert!(
-            (values[i] - i as f32).abs() < 1e-6,
-            "mismatch at index {i}: got {}",
-            values[i]
+            (v - i as f32).abs() < 1e-6,
+            "mismatch at index {i}: got {v}"
        );
    }
 }
@@ -1272,7 +1274,7 @@ fn write_roundtrip_scalar_f64_attr() {
    let mut fw = FileWriter::new();
    fw.create_dataset("data")
        .with_f64_data(&[1.0])
-        .set_attr("scale", AttrValue::F64(3.14));
+        .set_attr("scale", AttrValue::F64(3.25));
    let bytes = fw.finish().unwrap();

    let sig = find_signature(&bytes).unwrap();
@@ -1283,7 +1285,7 @@ fn write_roundtrip_scalar_f64_attr() {
    let scale = find_attribute(&attrs, "scale").expect("scale attr not found");
    let vals = scale.read_as_f64().unwrap();
    assert_eq!(vals.len(), 1);
-    assert!((vals[0] - 3.14).abs() < 1e-10);
+    assert!((vals[0] - 3.25).abs() < 1e-10);
 }

 #[test]
@@ -2,6 +2,15 @@

 use clawhdf5_format::data_read::{read_object_references, read_region_references};
 use clawhdf5_format::datatype::{Datatype, ReferenceType};
+/// The Python interpreter to drive interop checks with.
+///
+/// `CLAWHDF5_PYTHON` lets these run against a virtualenv holding h5py, which
+/// on a PEP 668 "externally managed" system is the only place it can be
+/// installed. Without it the suite silently skips, and a silent skip here is
+/// how a datatype bug once reached a release.
+fn python() -> String {
+    std::env::var("CLAWHDF5_PYTHON").unwrap_or_else(|_| "python3".to_string())
+}

 #[test]
 fn object_ref_single_valid() {
@@ -173,22 +182,27 @@ print('ok')
 "#,
        path.display()
    );
-    let output = std::process::Command::new("python3")
+    let output = std::process::Command::new(python())
        .args(["-c", &script])
        .output();

    let output = match output {
        Ok(o) if o.status.success() => o,
        _ => {
+            // CI sets CLAWHDF5_REQUIRE_INTEROP=1 so this can't silently skip.
+            assert!(
+                !std::env::var("CLAWHDF5_REQUIRE_INTEROP").is_ok_and(|v| v == "1"),
+                "CLAWHDF5_REQUIRE_INTEROP=1 but python3 with h5py is not available"
+            );
            eprintln!("skipping h5py_object_reference_roundtrip: python3+h5py not available");
            return;
        }
    };
    let stdout = String::from_utf8(output.stdout).unwrap();
-    if !stdout.trim().contains("ok") {
-        eprintln!("skipping h5py_object_reference_roundtrip: h5py script failed");
-        return;
-    }
+    assert!(
+        stdout.trim().contains("ok"),
+        "h5py reference-file generator did not report ok: {stdout}"
+    );

    // Read the file and parse object references
    let file_data = std::fs::read(&path).unwrap();
@@ -311,3 +325,97 @@ print('ok')
    // Clean up
    let _ = std::fs::remove_file(&path);
 }
+
+// ---------------------------------------------------------------------------
+// H5T_STD_REF (HDF5 1.12+ references, datatype message version 4)
+// ---------------------------------------------------------------------------
+
+/// `fixtures/std_ref_hdf5_2_0.h5` (see `gen_std_ref.py`) holds a dataset of
+/// `H5T_STD_REF` with two object references, written by HDF5 2.0 itself. The
+/// datatype used to be rejected with `InvalidReferenceType(2)`.
+#[test]
+fn std_ref_object_references_from_hdf5_2_0() {
+    use clawhdf5_format::data_layout::DataLayout;
+    use clawhdf5_format::dataspace::Dataspace;
+    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::superblock::Superblock;
+
+    let bytes: &[u8] = include_bytes!("fixtures/std_ref_hdf5_2_0.h5");
+    let sb = Superblock::parse(bytes, find_signature(bytes).unwrap()).unwrap();
+    let (os, ls) = (sb.offset_size, sb.length_size);
+
+    let refs_addr = resolve_path_any(bytes, &sb, "refs").unwrap();
+    let header = ObjectHeader::parse(bytes, refs_addr as usize, os, ls).unwrap();
+    let message = |t: MessageType| {
+        &header
+            .messages
+            .iter()
+            .find(|m| m.msg_type == t)
+            .unwrap()
+            .data
+    };
+
+    let (datatype, _) = Datatype::parse(message(MessageType::Datatype)).unwrap();
+    assert_eq!(
+        datatype,
+        Datatype::Reference {
+            size: 18,
+            ref_type: ReferenceType::Object2
+        }
+    );
+    let dataspace = Dataspace::parse(message(MessageType::Dataspace), ls).unwrap();
+    let layout = DataLayout::parse(message(MessageType::DataLayout), os, ls).unwrap();
+    let raw =
+        clawhdf5_format::data_read::read_raw_data(bytes, &layout, &dataspace, &datatype).unwrap();
+    assert_eq!(raw.len(), 2 * 18);
+
+    // The references point at the objects they were created from.
+    let refs = read_object_references(&raw, &datatype, os).unwrap();
+    let addresses: Vec<u64> = refs.iter().map(|r| r.address).collect();
+    assert_eq!(
+        addresses,
+        [
+            resolve_path_any(bytes, &sb, "target").unwrap(),
+            resolve_path_any(bytes, &sb, "grp").unwrap(),
+        ]
+    );
+    // And what they point at is a real object header.
+    for address in addresses {
+        ObjectHeader::parse(bytes, address as usize, os, ls).unwrap();
+    }
+}
+
+#[test]
+fn std_ref_decoding_rejects_malformed_elements() {
+    let dt = Datatype::Reference {
+        size: 18,
+        ref_type: ReferenceType::Object2,
+    };
+    let mut good = vec![0u8; 18];
+    good[..4].copy_from_slice(&[2, 0, 8, 0xb3]);
+    assert_eq!(
+        read_object_references(&good, &dt, 8).unwrap()[0].address,
+        0xb3
+    );
+
+    // Null reference.
+    assert_eq!(
+        read_object_references(&[0u8; 18], &dt, 8).unwrap()[0].address,
+        u64::MAX
+    );
+    for (what, patch) in [
+        ("wrong reference type", (0usize, 3u8)),
+        ("external flag", (1, 1)),
+        ("token longer than the element", (2, 200)),
+        ("zero-length token", (2, 0)),
+    ] {
+        let mut bad = good.clone();
+        bad[patch.0] = patch.1;
+        assert!(read_object_references(&bad, &dt, 8).is_err(), "{what}");
+    }
+    // Not a whole number of elements.
+    assert!(read_object_references(&good[..17], &dt, 8).is_err());
+}
@@ -4,9 +4,18 @@
 //! (and vice versa). They require python3 + h5py to be installed.

 use clawhdf5_format::file_writer::{AttrValue, CompoundTypeBuilder, EnumTypeBuilder, FileWriter};
+/// The Python interpreter to drive interop checks with.
+///
+/// `CLAWHDF5_PYTHON` lets these run against a virtualenv holding h5py, which
+/// on a PEP 668 "externally managed" system is the only place it can be
+/// installed. Without it the suite silently skips, and a silent skip here is
+/// how a datatype bug once reached a release.
+fn python() -> String {
+    std::env::var("CLAWHDF5_PYTHON").unwrap_or_else(|_| "python3".to_string())
+}

 fn h5py_available() -> bool {
-    std::process::Command::new("python3")
+    std::process::Command::new(python())
        .args(["-c", "import h5py"])
        .output()
        .map(|o| o.status.success())
@@ -17,10 +26,10 @@ fn h5py_read(_path: &std::path::Path, script: &str) -> String {
    if !h5py_available() {
        panic!("h5py not installed — skipping interop test");
    }
-    let o = std::process::Command::new("python3")
+    let o = std::process::Command::new(python())
        .args(["-c", script])
        .output()
-        .expect("python3");
+        .expect("python interpreter");
    if !o.status.success() {
        panic!("h5py: {}", String::from_utf8_lossy(&o.stderr));
    }
@@ -235,17 +244,31 @@ fn h5py_reads_our_array_dataset() {
 #[test]
 #[ignore = "requires Python h5py module"]
 fn read_h5py_generated_compound() {
-    let path = std::env::temp_dir().join("clawhdf5_h5py_compound.h5");
+    check_h5py_generated_compound("latest", ", libver='latest'");
+}
+
+/// Same file written with h5py's default format bounds. HDF5 2.0 raised the
+/// default low bound to 1.8, so "default" files exercise different on-disk
+/// structures than both `libver='latest'` and pre-2.0 defaults.
+#[test]
+#[ignore = "requires Python h5py module"]
+fn read_h5py_generated_compound_default_libver() {
+    check_h5py_generated_compound("default", "");
+}
+
+fn check_h5py_generated_compound(tag: &str, libver_kw: &str) {
+    let path = std::env::temp_dir().join(format!("clawhdf5_h5py_compound_{tag}.h5"));
    let gen_script = format!(
        r#"
 import h5py, numpy as np
 dt = np.dtype([('x', 'f8'), ('y', 'f8'), ('id', 'i4')])
 data = np.array([(1.0, 2.0, 10), (3.0, 4.0, 20)], dtype=dt)
-f = h5py.File('{}', 'w', libver='latest')
+f = h5py.File('{}', 'w'{})
 f.create_dataset('particles', data=data)
 f.close()
 "#,
-        path.display()
+        path.display(),
+        libver_kw
    );
    h5py_read(&path, &gen_script);

@@ -292,20 +315,102 @@ f.close()
    assert_eq!(x_vals, vec![1.0, 3.0]);
 }

+#[test]
+#[ignore = "requires Python h5py module"]
+fn read_h5py_generated_native_complex() {
+    // HDF5 2.0 native complex (datatype class 11, version 5), written through
+    // h5py's low-level API. Skips when the linked HDF5 predates 2.0.
+    let path = std::env::temp_dir().join("clawhdf5_h5py_native_complex.h5");
+    let gen_script = format!(
+        r#"
+import h5py, numpy as np
+from h5py import h5t, h5s, h5d, h5f, h5p
+if not getattr(h5py.get_config(), 'has_native_complex', False):
+    print('SKIP')
+else:
+    fapl = h5p.create(h5p.FILE_ACCESS)
+    fapl.set_libver_bounds(h5f.LIBVER_LATEST, h5f.LIBVER_LATEST)
+    fid = h5f.create(b'{}', h5f.ACC_TRUNC, fapl=fapl)
+    t = h5t.COMPLEX_IEEE_F64LE
+    d = h5d.create(fid, b'z', t, h5s.create_simple((2,)))
+    d.write(h5s.ALL, h5s.ALL, np.array([1+2j, 3+4j], dtype=np.complex128), mtype=t)
+    fid.close()
+"#,
+        path.display()
+    );
+    if h5py_read(&path, &gen_script) == "SKIP" {
+        eprintln!("HDF5 < 2.0: no native complex support, skipping");
+        return;
+    }
+
+    let bytes = std::fs::read(&path).unwrap();
+    let sig = clawhdf5_format::signature::find_signature(&bytes).unwrap();
+    let sb = clawhdf5_format::superblock::Superblock::parse(&bytes, sig).unwrap();
+    let addr = clawhdf5_format::group_v2::resolve_path_any(&bytes, &sb, "z").unwrap();
+    let hdr = clawhdf5_format::object_header::ObjectHeader::parse(
+        &bytes,
+        addr as usize,
+        sb.offset_size,
+        sb.length_size,
+    )
+    .unwrap();
+    let msg = |t: clawhdf5_format::message_type::MessageType| {
+        &hdr.messages.iter().find(|m| m.msg_type == t).unwrap().data
+    };
+    let (dt, _) = clawhdf5_format::datatype::Datatype::parse(msg(
+        clawhdf5_format::message_type::MessageType::Datatype,
+    ))
+    .unwrap();
+    let ds = clawhdf5_format::dataspace::Dataspace::parse(
+        msg(clawhdf5_format::message_type::MessageType::Dataspace),
+        sb.length_size,
+    )
+    .unwrap();
+    let dl = clawhdf5_format::data_layout::DataLayout::parse(
+        msg(clawhdf5_format::message_type::MessageType::DataLayout),
+        sb.offset_size,
+        sb.length_size,
+    )
+    .unwrap();
+    let raw = clawhdf5_format::data_read::read_raw_data(&bytes, &dl, &ds, &dt).unwrap();
+    let fields = clawhdf5_format::data_read::read_compound_fields(&raw, &dt).unwrap();
+    assert_eq!(fields.len(), 2);
+    let re =
+        clawhdf5_format::data_read::read_as_f64(&fields[0].raw_data, &fields[0].datatype).unwrap();
+    let im =
+        clawhdf5_format::data_read::read_as_f64(&fields[1].raw_data, &fields[1].datatype).unwrap();
+    assert_eq!((fields[0].name.as_str(), re), ("r", vec![1.0, 3.0]));
+    assert_eq!((fields[1].name.as_str(), im), ("i", vec![2.0, 4.0]));
+}
+
 #[test]
 #[ignore = "requires Python h5py module"]
 fn read_h5py_generated_enum() {
-    let path = std::env::temp_dir().join("clawhdf5_h5py_enum.h5");
+    check_h5py_generated_enum("latest", ", libver='latest'");
+}
+
+/// Same file written with h5py's default format bounds. HDF5 2.0 raised the
+/// default low bound to 1.8, so "default" files exercise different on-disk
+/// structures than both `libver='latest'` and pre-2.0 defaults.
+#[test]
+#[ignore = "requires Python h5py module"]
+fn read_h5py_generated_enum_default_libver() {
+    check_h5py_generated_enum("default", "");
+}
+
+fn check_h5py_generated_enum(tag: &str, libver_kw: &str) {
+    let path = std::env::temp_dir().join(format!("clawhdf5_h5py_enum_{tag}.h5"));
    let gen_script = format!(
        r#"
 import h5py, numpy as np
 dt = h5py.enum_dtype({{"RED": 0, "GREEN": 1, "BLUE": 2}}, basetype=np.int32)
 data = np.array([1, 0, 2, 1], dtype=np.int32)
-f = h5py.File('{}', 'w', libver='latest')
+f = h5py.File('{}', 'w'{})
 f.create_dataset('colors', data=data, dtype=dt)
 f.close()
 "#,
-        path.display()
+        path.display(),
+        libver_kw
    );
    h5py_read(&path, &gen_script);

@@ -1,10 +1,10 @@
 [package]
 name = "clawhdf5-gpu"
-version = "2.1.0"
+version = "2.6.0"
 edition = "2024"
 description = "GPU-accelerated vector operations for rustyhdf5 using wgpu compute shaders"
 license = "MIT"
-repository = "https://github.com/redclawsystems/clawhdf5"
+repository = "https://git.redclaw.dev/quantumclaw/clawhdf5"
 readme = "README.md"
 keywords = ["hdf5", "gpu", "wgpu", "compute"]
 categories = ["science", "graphics"]
@@ -6,6 +6,9 @@ use crate::shaders;
 use bytemuck::Pod;
 use wgpu::util::DeviceExt;

+/// Upper bound on a single GPU→CPU readback wait.
+const READBACK_TIMEOUT: std::time::Duration = std::time::Duration::from_secs(30);
+
 /// GPU-accelerated vector search engine.
 ///
 /// Upload vectors once, then run many searches against them.
@@ -1033,10 +1036,12 @@ impl GpuAccelerator {
        slice.map_async(wgpu::MapMode::Read, move |result| {
            let _ = tx.send(result);
        });
+        // Bounded wait: a wedged driver must surface as an error, not hang
+        // the caller forever.
        self.device
            .poll(wgpu::PollType::Wait {
                submission_index: None,
-                timeout: None,
+                timeout: Some(READBACK_TIMEOUT),
            })
            .map_err(|e| GpuError::BufferMap(format!("device poll failed: {e}")))?;
        rx.recv()
@@ -6,9 +6,41 @@
 mod tests {
    use clawhdf5_gpu::{GpuAccelerator, GpuError};

-    fn skip_if_no_gpu() -> Option<GpuAccelerator> {
+    /// Serialises GPU access across tests. The harness runs tests on many
+    /// threads; letting each create its own wgpu instance + device (with
+    /// adapter-maximum limits) at the same time can wedge the driver and hang
+    /// the whole suite, so every test holds this lock while it owns a device.
+    static GPU_LOCK: std::sync::Mutex<()> = std::sync::Mutex::new(());
+
+    fn gpu_lock() -> std::sync::MutexGuard<'static, ()> {
+        // A panicking test poisons the lock; the guarded state is `()`.
+        GPU_LOCK.lock().unwrap_or_else(|e| e.into_inner())
+    }
+
+    /// A `GpuAccelerator` plus the lock that keeps other tests off the GPU.
+    /// Field order matters: the device is dropped before the lock is released.
+    struct LockedGpu {
+        gpu: GpuAccelerator,
+        _guard: std::sync::MutexGuard<'static, ()>,
+    }
+
+    impl std::ops::Deref for LockedGpu {
+        type Target = GpuAccelerator;
+        fn deref(&self) -> &GpuAccelerator {
+            &self.gpu
+        }
+    }
+
+    impl std::ops::DerefMut for LockedGpu {
+        fn deref_mut(&mut self) -> &mut GpuAccelerator {
+            &mut self.gpu
+        }
+    }
+
+    fn skip_if_no_gpu() -> Option<LockedGpu> {
+        let guard = gpu_lock();
        match GpuAccelerator::new() {
-            Ok(gpu) => Some(gpu),
+            Ok(gpu) => Some(LockedGpu { gpu, _guard: guard }),
            Err(_) => {
                eprintln!("SKIPPED: no GPU available");
                None
@@ -69,6 +101,7 @@ mod tests {
    #[test]
    fn test_gpu_availability_detection() {
        // Should not panic regardless of GPU presence
+        let _guard = gpu_lock();
        let available = GpuAccelerator::is_available();
        eprintln!("GPU available: {available}");
    }
@@ -425,6 +458,7 @@ mod tests {
    #[test]
    fn test_graceful_no_gpu_fallback() {
        // This test just demonstrates the pattern — it always passes
+        let _guard = gpu_lock();
        match GpuAccelerator::new() {
            Ok(gpu) => {
                eprintln!("GPU found: {}", gpu.device_info());
@@ -1,16 +1,16 @@
 [package]
 name = "clawhdf5-io"
-version = "2.1.0"
+version = "2.6.0"
 edition = "2024"
 description = "I/O abstraction layer for rustyhdf5"
 license = "MIT"
-repository = "https://github.com/redclawsystems/clawhdf5"
+repository = "https://git.redclaw.dev/quantumclaw/clawhdf5"
 readme = "README.md"
 keywords = ["hdf5", "io", "science", "data"]
 categories = ["filesystem", "science"]

 [dependencies]
-clawhdf5-format = { path = "../clawhdf5-format", version = "2.1.0" }
+clawhdf5-format = { path = "../clawhdf5-format", version = "2.6.0" }
 memmap2 = { version = "0.9", optional = true }
 libc = { version = "0.2", optional = true }
 tokio = { version = "1", features = ["fs", "io-util"], optional = true }
@@ -59,11 +59,16 @@ pub trait AsyncHDF5Read: Send + Sync {

 /// Async file-backed reader using tokio for non-blocking I/O.
 ///
-/// Opens a file and reads it asynchronously. The file is read into memory
-/// on first access, making subsequent operations fast.
+/// Opens a file and reads it asynchronously. The underlying file handle is
+/// opened once (lazily, on first access) and cached for the lifetime of this
+/// reader, so repeated granular `read_at` calls reuse the open descriptor
+/// and cached length instead of paying an open+stat syscall pair every time.
+/// The handle is guarded by a mutex, which also correctly serializes the
+/// seek-then-read pairs of concurrent callers sharing the one file position.
 #[derive(Debug)]
 pub struct AsyncFileReader {
    path: std::path::PathBuf,
+    handle: tokio::sync::Mutex<Option<(tokio::fs::File, u64)>>,
 }

 impl AsyncFileReader {
@@ -73,6 +78,7 @@ impl AsyncFileReader {
    pub fn new<P: AsRef<Path>>(path: P) -> Self {
        Self {
            path: path.as_ref().to_path_buf(),
+            handle: tokio::sync::Mutex::new(None),
        }
    }

@@ -89,23 +95,33 @@ impl AsyncFileReader {

 impl AsyncHDF5Read for AsyncFileReader {
    async fn read_at(&self, offset: u64, len: usize) -> io::Result<Vec<u8>> {
-        let mut file = tokio::fs::File::open(&self.path).await?;
-        let metadata = file.metadata().await?;
-        let file_len = metadata.len();
+        let mut guard = self.handle.lock().await;
+        if guard.is_none() {
+            let file = tokio::fs::File::open(&self.path).await?;
+            let file_len = file.metadata().await?.len();
+            *guard = Some((file, file_len));
+        }
+        let (file, file_len) = guard.as_mut().expect("just populated above");
+        let file_len = *file_len;
        if offset >= file_len {
            return Ok(Vec::new());
        }
        let available = (file_len - offset) as usize;
        let to_read = len.min(available);
-        tokio::io::AsyncSeekExt::seek(&mut file, io::SeekFrom::Start(offset)).await?;
+        tokio::io::AsyncSeekExt::seek(file, io::SeekFrom::Start(offset)).await?;
        let mut buf = vec![0u8; to_read];
        file.read_exact(&mut buf).await?;
        Ok(buf)
    }

    async fn len(&self) -> io::Result<u64> {
-        let metadata = tokio::fs::metadata(&self.path).await?;
-        Ok(metadata.len())
+        let mut guard = self.handle.lock().await;
+        if guard.is_none() {
+            let file = tokio::fs::File::open(&self.path).await?;
+            let file_len = file.metadata().await?.len();
+            *guard = Some((file, file_len));
+        }
+        Ok(guard.as_ref().expect("just populated above").1)
    }
 }

@@ -1,10 +1,10 @@
 [package]
 name = "clawhdf5-migrate"
-version = "2.1.0"
+version = "2.6.0"
 edition = "2024"
 description = "CLI to migrate SQLite agent memory databases to HDF5 format"
 license = "MIT"
-repository = "https://github.com/redclawsystems/clawhdf5"
+repository = "https://git.redclaw.dev/quantumclaw/clawhdf5"
 readme = "README.md"
 keywords = ["sqlite", "hdf5", "migration", "agent", "memory"]
 categories = ["command-line-utilities", "database"]
@@ -14,9 +14,9 @@ name = "clawhdf5-migrate"
 path = "src/main.rs"

 [dependencies]
-clawhdf5-agent = { path = "../clawhdf5-agent", version = "2.1.0" }
-clawhdf5-format = { path = "../clawhdf5-format", version = "2.1.0" }
-clawhdf5 = { path = "../clawhdf5", version = "2.1.0" }
+clawhdf5-agent = { path = "../clawhdf5-agent", version = "2.6.0" }
+clawhdf5-format = { path = "../clawhdf5-format", version = "2.6.0" }
+clawhdf5 = { path = "../clawhdf5", version = "2.6.0" }
 rusqlite = { version = "0.31", features = ["bundled"] }
 clap = { version = "4", features = ["derive"] }
 half = { workspace = true }
@@ -49,6 +49,10 @@ pub fn read_hdf5(path: &str) -> Result<SqliteData, BoxErr> {
        entities,
        relations,
        embedding_dim,
+        // Not a SQLite read — the caller (incremental migration) carries
+        // forward the current run's actual `source_path` from the fresh
+        // SQLite read instead of using this placeholder.
+        source_path: String::new(),
    })
 }

--- a/Show More
+++ b/Show More