search() collected ef candidates, then filtered out soft-deleted nodes, then
took k. When the records nearest a query had been deleted, every candidate was
a tombstone and the search returned fewer than k results — 39 of 40 queries in
the new test, which deletes each query's 40 nearest neighbours.
search_layer takes an optional skip mask: a skipped node is still pushed onto
the candidate queue (a tombstone is a valid waypoint) but never into the
result heap, so the ef result slots hold live nodes only. Build and insert
pass no mask. Recall and speed without deletions are unchanged.
Co-Authored-By: Claude Fable 5.1 <[email protected]>
Bump all workspace crates, the node package and pyproject to 2.4.0, finalize
the changelog and add upgrade notes for the search behaviour changes.
Co-Authored-By: Claude Fable 5.1 <[email protected]>
Fusion min-max normalises over every keyword match, so hybrid_search asked
BM25 for a ranked list of the whole corpus: a hash insert per posting, then a
sort of every match, then the merge sorted every candidate again to keep k.
- BM25Index::scores returns every match unsorted, accumulated in a dense array
(contributions are strictly positive, so zero means untouched). search() is
built on it with the bounded heap.
- merge_vector_keyword partitions out its top k (select_nth) and orders only
those, with the same score-then-id order.
- Both hybrid paths use scores().
Rankings are identical (equivalence tests for both changes). p50 0.24 -> 0.07
ms (1K), 2.1 -> 0.49 ms (10K), 23 -> 4.65 ms (100K).
The harness gains --fusion-study, which measured the alternative — capping the
keyword pool — and found it changes the top-10 for most queries (overlap
0.83-0.92, different #1 for 10-35%) for only a 2x saving. Not adopted.
Co-Authored-By: Claude Fable 5.1 <[email protected]>
- Cosine distance was 1 - dot/(|a||b|), re-deriving both norms on every
evaluation in the innermost loop of build and search. The index now stores
unit vectors (prepared at build, insert, graph load and HDF5 load; the query
once per search) and uses 1 - dot. Zero vectors stay zero, giving distance 1
as before. Returned distances are unchanged.
- search_layer allocated a HashSet of visited nodes per call. It is now an
epoch-stamped u32 array in thread-local scratch, reused across calls, so
search(&self) stays shareable between threads.
- clawhdf5-accel caches the detected SIMD backend in a OnceLock.
Recall is identical. Build 2.75 -> 1.89 s (10K), ~38 -> 21 s (100K); QPS at
ef=64 22.7K -> 39K (10K), 10.4K -> 14K (100K).
Co-Authored-By: Claude Fable 5.1 <[email protected]>
open() marked the HNSW index dirty, so the first search of every session
rebuilt it from scratch — 36 s at 100K records with the (better, slower)
heuristic build. First query after open is now 1.7 / 15 / 159 ms at
1K / 10K / 100K; what remains is the one-off keyword index build.
- clawhdf5-ann: HnswIndex::graph_to_bytes / from_graph_bytes serialize the
graph only (levels, tombstones, adjacency as u32, CRC32). The existing HDF5
serializer embeds a full copy of every vector, which would double a store
that already holds them. Loading validates everything — counts, levels vs
layer count, connection limits, every neighbour id and the layer it must
exist on — so a damaged graph, or a hostile one with a valid checksum, is an
error rather than an out-of-bounds walk during search.
- clawhdf5-agent: each checkpoint writes the graph to <store>.h5.ann (synced,
atomic, before the .h5) and records a fresh generation id in /meta. open()
loads the sidecar only if its generation matches that checkpoint; missing,
stale, damaged or mismatched sidecars are ignored and the index rebuilt.
Records appended through WAL replay join the loaded index incrementally; a
replayed Update or Tombstone invalidates it. snapshot() copies it. Only an
index that exactly mirrors the cache is saved; otherwise a stale sidecar is
removed.
- ensure_hnsw_fresh inserts records appended since the last sync instead of
rebuilding, so save_batch no longer marks the whole index dirty.
- CheckpointMeta { wal_applied, ann_generation } with *_with_meta build/write/
read functions; the *_with_mark ones delegate.
- Harness reports the one-off cold index build separately from the first query
after a reopen.
Co-Authored-By: Claude Fable 5.1 <[email protected]>
hybrid_search rebuilt the BM25 index from scratch (re-tokenising every record)
and rewrote the whole .h5 file on every single query, so a query cost O(store
size) in both CPU and disk I/O. Steady-state p50 per the search harness:
5.5 -> 0.24 ms (1K), 49 -> 2.1 ms (10K), 884 -> 23 ms (100K).
- BM25Index is incremental: add_document / remove_document keep it exactly
equivalent to a fresh build over the same live documents (property test: 60
random op sequences compared against BM25Index::build after every step). IDF
moves to query time since it depends on the live document count. Top-k uses
a bounded heap, ties break by doc id (results were HashMap-ordered), and the
"WAND" code that computed a bound and then discarded it is removed.
- HDF5Memory keeps one index for its lifetime, built lazily. Appends are
picked up by ensure_bm25_fresh whatever path added them; delete and in-place
update report themselves; compaction drops the index. A test drives every
mutation and compares against a fresh build.
- A query no longer calls flush(). Activation boosts are marked dirty and
persisted by the next checkpoint, including a best-effort one on drop so a
search-only session keeps them (approved behaviour change). Activation
weights are capped at 16.0; they previously grew without bound.
The archived mission branch's BM25 cache was reviewed and not used: it was
invalidated by every write, so interleaved save/search still rebuilt per
query, and it changed the default fusion weights.
Co-Authored-By: Claude Fable 5.1 <[email protected]>
Neighbours were selected as the plain closest-M, for both a new node's links
and back-link pruning. On clustered data every link of a node inside a tight
cluster then goes to that same cluster, so clusters become islands a search
entering elsewhere can never reach: recall@10 was 0.87 / 0.67 / 0.31 at
1K / 10K / 100K (384-dim) and flat in ef. Uniform random data — all the
existing tests used — does not show it.
Implement the HNSW paper's Algorithm 4 with keepPrunedConnections: accept a
candidate only if it is closer to the node than to every neighbour already
accepted, then fill spare slots with the closest rejected ones. Recall@10 at
ef=64 is now 1.00 / 1.00 / 0.98 and rises with ef; uniform data improves
slightly. Build is ~3.5x slower at 10K (extra distance evaluations), to be
recovered by the distance-kernel work. The needless rayon fan-out over <=33
distances in prune_connections is gone.
Tests: a clustered-data recall test for bulk build and incremental insert
(scores 0.43 with the old selection), and a unit test of the selection rule.
Harness gains --uniform and --ann-only; before/after in BENCHMARKS.md.
Co-Authored-By: Claude Fable 5.1 <[email protected]>
Bump all workspace crates, the node package and pyproject to 2.3.0, finalize
the changelog and add upgrade notes for the behaviour changes.
Co-Authored-By: Claude Fable 5.1 <[email protected]>
attrs() silently omitted any attribute whose datatype had no AttrValue variant
— including every Python bool, which h5py stores as an enum — plus complex,
compound and reference attributes, and cast unsigned 64-bit arrays to
I64Array so values above i64::MAX came back negative.
- numpy/h5py-style booleans (an enum of exactly FALSE=0 / TRUE=1 over an
integer base) decode as I64 / I64Array of 0/1.
- AttrValue::U64Array keeps unsigned arrays unsigned. Behaviour change: an
unsigned array attribute no longer arrives as I64Array; the netCDF-4 CF
helpers (_FillValue, valid_range) and the Python bindings handle it.
- AttrValue::Raw { datatype, shape, data } carries any other attribute
verbatim (also used when a value fails to decode as its declared type), so
the attribute list is always complete. Decodable with data_read against the
datatype; Python receives {"dtype", "shape", "data"}.
- Both new variants are writable, so attributes round-trip between files.
h5py interop tests cover reading 13 attribute kinds and h5py reading back a
compound and a u64 attribute written by clawhdf5.
Co-Authored-By: Claude Fable 5.1 <[email protected]>
- clippy --all-targets plus a clawhdf5-format feature matrix (parallel, lz4,
zstd, pcodec, fast-checksum); fix the accumulated lint backlog in test,
bench and feature-gated code (no behaviour changes).
- Install python3 + h5py/numpy/netCDF4/xarray in the CI container and set
CLAWHDF5_REQUIRE_INTEROP=1, which makes a missing interop dependency a test
failure. Every h5py/netCDF4 interop test used to skip silently in CI. Run
the #[ignore]d writer_h5py_tests suite explicitly.
- cargo bench --no-run so benches can't rot; fix bench.rs and memory_bench.rs,
which no longer compiled against the current strategy/consolidation APIs.
- Optional fuzz smoke run via CLAWHDF5_FUZZ_SECONDS.
- CHANGELOG and docs/known-issues.md updated.
Co-Authored-By: Claude Fable 5.1 <[email protected]>
Bump all workspace crates, the node package and pyproject to 2.2.0 and
finalize the changelog. The repository URL in every manifest pointed at a
GitHub location that does not resolve; use the real origin.
Co-Authored-By: Claude Fable 5.1 <[email protected]>
Class 11 (datatype version 5) properties are a single base floating-point
datatype message, not a compound-style member list. The old parser read the
base type's bytes as member names, yielding a garbage datatype, and failed
with UnexpectedEof when a complex type was nested in a compound.
Parse the base type and surface the type as the equivalent {r, i} compound
(the shape h5py writes for numpy complex dtypes), with a size check against
the base type. Covered by byte-level tests taken from HDF5 2.0 output and an
h5py end-to-end test (writer_h5py_tests is now 27/27 against HDF5 2.0.0).
Found while validating a user report of InvalidDatatypeVersion
{ class: 6, version: 5 } against v2.1.0 (already fixed on main in a13ff51,
never released). Add docs/known-issues.md recording that report, this bug,
the open reference-v4 gap and a gpu_tests parallel-run hang; credit the
reporter in the changelog.
Co-Authored-By: Claude Fable 5.1 <[email protected]>
README.md:
- Fix badly stale LongMemEval numbers (badge said Hit@5 46%, table showed
fabricated ~46%/~0.34/~72% figures that never matched BENCHMARKS.md's
actual results of Hit@5 100% session / 84.4% turn-level, MRR 1.0/0.6597)
- Remove clawhdf5-types from the Crate Map — that crate was removed in an
earlier cleanup pass but the README diagram was never updated; fix the
crate count (16, not 17) and stale line-of-code figures (72,087/84K -> ~92K)
- Fix a dead #benchmarks badge anchor (no such heading exists) -> #performance
- Document the new clawhdf5-ann `parallel` feature (had no Feature Flags entry)
- Note WAL's CRC32 per-entry check, link the new tank LongMemEval/SIMD/
vector-search reproduction section, update stale test-count comment
(417+ -> 1,650+) and Phase 2 roadmap blurb (LongMemEval is now done)
ROADMAP.md:
- Check off "Academic benchmark cross-validation" (done via the tank
LongMemEval re-run) and add a new "Recently closed out" section
summarizing the Tier 3-4 hardening pass (Android JNI validation, pyo3
bump, WAL CRC32, bounds-check audit + fuzz harness that found 3 real
bugs, HNSW optional parallel feature, workspace.dependencies)
- Update stale test count (1,546 -> 1,650+) and last-updated date
CLAUDE.md: mention WAL's per-entry CRC32 check
CHANGELOG.md: add Security/Performance/Architecture/Documentation entries
under Unreleased summarizing all of Tiers 1-4 (this had not been touched
since 2026-06-04, predating the entire hardening pass)
clawhdf5-migrate:
- Real content validation: the post-migration check reads the written HDF5
back (new hdf5_reader) and compares actual content — chunk text, embeddings,
and every session/entity/relation field — to the source, not just row counts.
A representative sample of chunk rows is verified by default; --validate-full
checks every row. A count-preserving corruption no longer passes.
- Configurable schema: SQL is built from a SchemaConfig (table + ordered column
names, defaulting to the ZeroClaw layout) instead of hardcoded queries, with
--chunks-table / --sessions-table / --entities-table / --relations-table.
- Streaming count pass: --dry-run does a COUNT(*)-only pass per table instead
of loading every row.
- Incremental migration: --incremental reads the existing output, reads only
source chunks with id greater than the last migrated id, and appends them
(metadata groups refreshed from source) rather than re-migrating everything.
clawhdf5-format:
- read_as_f32 / read_as_f64 now decode IEEE-754 half-precision (2-byte) floats
via a no_std-safe bit conversion — needed to read float16-stored embeddings
back (e.g. for migrate's content validation), previously a TypeMismatch.
Tests: f16 read unit test; migrate tests for content-corruption detection,
custom table names, and incremental append; CLI smoke-tested end-to-end and the
dense/incremental output verified with h5py.
Co-Authored-By: Claude Opus 4.8 <[email protected]>
Dense attribute and dense link storage capped at a single fractal-heap direct
block (~64 KiB of heap data — a few thousand objects); beyond that the writer
produced an invalid oversized block. Lift the cap with a root indirect block.
When the serialized objects don't fit in one direct block, build a root
indirect block (FHIB) over multiple direct blocks sized by the doubling table
(start 512, width 4, doubling per row up to 64 KiB). Objects are packed
row-major across blocks, each block carries its logical block offset, and heap
IDs encode each object's heap offset (block offset + position). The FRHP points
root -> FHIB with the row count; unused slots in the current rows are undefined.
The fractal-heap builder is unified: FractalHeapBlock now carries the full heap
blob, and a shared write_frhp helper serializes the header for both the
single-block and multi-block paths. The single-block path is unchanged
(byte-identical), so existing dense attrs/links stay valid.
Validated end-to-end: a 2,500-attribute object and a 2,500-link group round-trip
through our reader and are read correctly by h5py. Objects still may not span a
block (no huge-object path).
Tests: facade round-trips for multi-block dense attrs and dense links, plus an
h5py-gated interop test (verified against the real h5py environment).
Co-Authored-By: Claude Opus 4.8 <[email protected]>
A group with more than 8 links (libhdf5's compact max_compact default) is now
written densely instead of as inline Link messages: the links live in a
single-direct-block fractal heap indexed by a v2 B-tree of type 5 (link-name
index), referenced from the group's LinkInfo message. This matches libhdf5's
compact->dense switchover and keeps large groups out of the object header.
- Extract the byte-identical fractal-heap builder shared by dense attributes
and dense links, parameterized by heap_id_length / max_heap_size. Attributes
keep 8 / 40; links use 7 / 32 to match libhdf5 (reverse-engineered: an
h5py-written dense group uses heap_id_length 7, max_heap_size 32, type-5
record = hash(4) + heap_id(7) = 11 bytes). This was the cause of an initial
"object overruns end of direct block" error from h5py.
- build_group_oh gained an optional dense LinkInfo (omitting inline Link
messages); the two-pass file assembly allocates each group's link blob after
its object header and rebuilds it with real target addresses in the final
pass (link-message size is address-independent, so layout is stable).
Validated end-to-end: our reader round-trips dense groups, h5py reads the
groups we write, and dense attributes remain byte-identical (still h5py-valid).
The agent's 9-dataset memory group now writes densely and round-trips. Single
direct block only (~a couple thousand links); indirect blocks remain a TODO.
Tests: facade round-trip (20-link dense group + compact sibling) and an
h5py-gated interop test confirming libhdf5 reads our dense groups.
Co-Authored-By: Claude Opus 4.8 <[email protected]>
The readers added this cycle parse untrusted bytes, so malformed/hostile
input must produce errors — never a panic, OOM, or unbounded recursion.
Audited each new surface and fixed the concrete vectors, each covered by an
adversarial regression test:
- Paged Fixed Array: `1 << max_nelmts_bits` shift overflow (u8 up to 255);
element-count bounded by file size; element/page offset multiplies checked.
- H5S selection decoder: ALL/NONE validate they have the 16 bytes they claim
to consume; hyperslab rank capped at 32 (H5S_MAX_RANK); iter_linear
coordinate/stride/product arithmetic uses checked ops.
- VDS mapping parser: drop pre-allocation from the untrusted `nused`;
bounds-check all selection slicing.
- scale-offset / N-Bit filters: `1 << minbits` overflow at minbits==64; N-Bit
`bit_offset + precision` overflow; N-Bit type-tree recursion depth capped to
stop a crafted nested tree from overflowing the stack; element counts bounded
by the chunk's expected decompressed size (threaded the previously-unused
chunk_size into both decoders) so a bogus count can't over-allocate.
- VDS assembly: a virtual dataset whose source is itself virtual (a cycle) now
errors instead of recursing into a stack overflow.
16 new adversarial tests; full format suite (482 lib) + agent + facade green;
clippy clean.
Co-Authored-By: Claude Opus 4.8 <[email protected]>
clawhdf5-agent: fixed-length string datasets (memory text chunks, session
summaries, ids, tags, entity/relation names) were stored uncompressed behind
a stale "chunked compound not yet supported" comment. Chunked writes work for
fixed-size string/compound datatypes like any other, so write_string_dataset
now chunks + deflates once a dataset's payload reaches 4 KiB — large,
redundant NullPad content compresses well while tiny metadata stays
contiguous (no chunk-overhead bloat). The dead `compress` parameter is
removed in favor of this size heuristic.
clawhdf5-format: enabling string compression exposed a latent bug — the
per-file ChunkCache built its chunk index once and reused it for every
chunked dataset in the file, keyed only by chunk coordinate with no dataset
discrimination. With one chunked dataset per file this never surfaced; with
two of different rank (a 1-D compressed string array and the 2-D embeddings
matrix) the first dataset's rank-1 index was reused for the second, panicking
with an out-of-bounds chunk coordinate. The cache now binds to a dataset by
its chunk-index address and rebinds — dropping the index, chunk-index map,
layout, and decompressed slots — whenever the dataset being read changes,
while still caching repeated/sequential access to the same dataset.
Tests: facade regression reading a 1-D compressed string dataset and a 2-D
compressed f32 dataset through one shared File cache (verified to panic
without the fix); existing agent e2e tests (large text chunks, migration
round-trip) now pass with compression on.
Co-Authored-By: Claude Opus 4.8 <[email protected]>
The HDF5 library does not implement the scale-offset filter's floating-point
E-scale mode. When asked for it (cd_values[0] = 1) it stores the chunk raw
(no minbits/minval header) and sets the chunk filter mask to skip the filter,
so such datasets read back verbatim purely by honoring the per-chunk filter
mask — no E-scale decoder is required.
Add a fixture written via the HDF5 low-level API (exact-representable values)
and a test asserting it reads back verbatim, locking in the filter-mask path.
Co-Authored-By: Claude Opus 4.8 <[email protected]>
VDS sources living in other files were previously unsupported because the
pure-byte read API has no filesystem. Add a resolver seam and wire a default.
clawhdf5-format:
- Add VdsSourceResolver (Fn(&str) -> Option<Vec<u8>>) and
read_raw_data_full_with_resolver. read_virtual_data uses the resolver to
fetch an external source file's bytes by its stored name, then reads the
named source dataset from those bytes and scatters as usual. A resolver
returning None leaves the region at fill (HDF5's missing-source behavior);
an external source with no resolver at all is a clean error. read_raw_data_full
is unchanged (delegates with no resolver).
clawhdf5:
- File now records the directory it was opened from and, for virtual layouts,
reads through a default resolver that loads sibling source files relative to
that directory. So File::open(virt).dataset(d).read_*() transparently
assembles cross-file VDS. In-memory files (from_bytes) have no directory, so
only same-file VDS resolves there.
Tests: format-layer external read with an injected resolver (and the
no-resolver error path), plus facade tests that drop both files in a temp dir
and read through File::open — covering successful resolution and the
missing-source-is-fill case.
Co-Authored-By: Claude Opus 4.8 <[email protected]>
Generalize Selection iteration from 1-D to arbitrary rank: iter_linear(dims)
enumerates a selection's row-major linear indices over a dataspace of the
given shape (ALL, NONE, regular hyperslabs, points), which is the order HDF5
uses to pair virtual and source selections.
read_virtual_data now passes the full virtual/source dimensions instead of a
single extent, so multi-dimensional block mappings scatter to the correct
non-contiguous linear positions. read_named_dataset_raw returns the source
dataset's dimensions. The rank-1 restriction is removed; only external-file
sources remain unsupported.
Tests: 2-D integration fixture (vds_2d_same_file.h5: two 2x2 sources placed
as non-contiguous blocks in a 4x4 virtual) plus N-D iter_linear unit tests
(block, strided, ALL, rank-mismatch). The 1-D path is unchanged.
Co-Authored-By: Claude Opus 4.8 <[email protected]>
A virtual layout previously returned UnsupportedVersion. Implement reading
for the common 1-D, same-file case, reverse-engineered and validated against
HDF5 2.0.
- Rewrite parse_vds_mappings to the real global-heap block format
(version(1) · nused(length_size) · entries · checksum(4)), where each
entry is source-file(null) · source-dataset(null) · source-selection ·
virtual-selection. Block version 1 encodes a same-file source as a single
0x04 marker in place of the file name; version 0 stores an explicit file
name. The selections are H5S-serialized and self-describing in length, so
they are decoded to find entry boundaries. The previous parser used a
guessed layout that did not match real files.
- Extend Selection with decode_serialized() (H5S_select_serialize: ALL,
NONE, and version-3 regular hyperslabs) and iter_linear_1d().
- Add read_virtual_data: resolve the mapping block from the global heap,
read each same-file source dataset, and scatter its selected elements into
the virtual buffer; unmapped regions stay at the zero fill value.
External-file sources and N-D selections return a clean unsupported error.
Tests: real-file integration test (vds_same_file.h5: partial source slice +
fill gap), selection decoder unit tests built from the fixture bytes, and
same-file/external mapping-parser unit tests.
Co-Authored-By: Claude Opus 4.8 <[email protected]>
A filtered, fixed-dimension dataset with more than one Fixed Array
data-block page (>1024 chunks by default) previously failed with
"paged Fixed Array data blocks not yet supported".
Implement the paged data-block layout, reverse-engineered and validated
against an HDF5 2.0 file:
- after the FADB prefix: a page-init bitmap (one bit per page, MSB-first
within each byte), a 4-byte checksum, then the pages;
- each page is a fixed full-size slot of page_nelmts elements plus a
4-byte checksum, with only the final page shorter;
- uninitialized pages still occupy their slot (zero-filled), so the
bitmap — not a 0xFF sentinel — marks a whole page unallocated.
Element parsing is factored into parse_fa_element, shared by the
non-paged and paged paths.
Tests: real-file integration test against a minimal 2-page gzip fixture
(v4_fixed_array_paged.h5) plus a synthetic unit test covering a
multi-byte/MSB-first bitmap, a skipped uninitialized page, and a short
final page.
Co-Authored-By: Claude Opus 4.8 <[email protected]>
Records the parallel chunk-compression perf change and the doc sweep that
landed after the v2.1.0 tag.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Fold the Unreleased HNSW / Python 3.14 entries into a single v2.1.0
(2026-06-03) section being cut and tagged now.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
