# Release Qualification
## 0.1.8
`a3s-vec` `0.1.8` is the Rust 1.75 fix for the published `0.1.7` crate.
Tag `0.1.8` points at the revision that `cargo publish` uploads. The registry
SHA-256 is filled in the post-publish checklist below once that upload exists.
### 0.1.8 release notes
`f32::next_up` is unstable on Rust 1.75. One Flat rejector test used it, so
`cargo test` failed the MSRV job for `0.1.7`. The test now steps one ULP
toward `+∞` with `to_bits`. The library, the public `f64` contract, and the
2026-09-23 proof are unchanged.
### 0.1.8 post-publish checklist
1. Hosted CI on the tagged revision is green.
2. The published crate SHA-256 is recorded here.
3. Formal git tag `0.1.8` points at that revision, and `cargo publish`
uploaded the matching crate.
## 0.1.7
`a3s-vec` `0.1.7` is published. Tag `0.1.7` @
`57fc47694befd0f9915df21175a2158918432170`, crates.io SHA-256
`90254cfdaddbe3848788fd93fc91a301ac39016b60b1e09c6afca799d59efd73`.
`cargo test` on Rust 1.75 fails in `src/index/vector_query.rs` because that
revision calls `f32::next_up`. `0.1.8` replaces that call. The library itself
matches the ranking code measured on 2026-09-23. `0.1.6` stays on the
registry (tag `0.1.6` @ `c7b828cfa610bb3e1aca84ae9255f1caf7c938c7`, SHA-256
`67c238a010add5d35d085a87a9183b9d3e08aef6a89b1b5edf19186b31b65062`).
### 0.1.7 release notes
The document snapshot and the WAL remain authoritative. Derived indexes still
propose candidates, and the public score is still the exact `f64` re-rank.
Default HNSW `ef` stays 64, exact re-rank stays on, default durability stays
`Always`, and IVF queries gain no default `scale_factor`. HNSW construction
for a fixed corpus and seed still publishes the ordered single-thread `f64`
neighbor sets. Equal scores keep the ascending primary key, including a
zero-norm cosine top-k after the flat index is rebuilt.
- Filter parsing, tokenization, and FP16/INT8/INT4 quantization are Rust in
this crate. The optional `jieba` feature uses `jieba-rs`. The previous
private kernel dependency is gone.
- New WAL frames are version 5 packed MessagePack. Older JSON frames still
decode. Schema-only operations still require frame version 4 or newer.
- Accounted bytes grow with the entries written in a batch. A drift or
overflow falls back to a full measure.
- The 2026-09-23 same-host protocol (cosine, top-10, `m=16`,
`ef_construction=96`, `ef=64`, one worker, zvec 0.7.0 with the refiner off)
is the proof in [README.md](README.md) and [BENCHMARKS.md](BENCHMARKS.md).
The 2026-09-20 million-document insert of `77,339.081` ms stays an unsplit
historical measurement.
### 0.1.7 post-publish checklist
1. Hosted CI run `35818658069` on revision `57fc476` failed the Rust 1.75
MSRV job. `0.1.8` is the fix.
2. The published crate SHA-256 matches
`90254cfdaddbe3848788fd93fc91a301ac39016b60b1e09c6afca799d59efd73`.
3. Formal git tag `0.1.7` points at that revision, and `cargo publish`
uploaded the matching crate.
## 0.1.6 historical binding
`a3s-vec` `0.1.6` is published. Tag `0.1.6`, and crates.io SHA-256
`67c238a010add5d35d085a87a9183b9d3e08aef6a89b1b5edf19186b31b65062` bind to
revision `c7b828cfa610bb3e1aca84ae9255f1caf7c938c7`. Hosted CI run
`35754487571` is green for that revision. The prior `0.1.5` package remains
on the registry (tag `0.1.5` @ `84f8985a669d0e4d33eb0c4884c7ccb90e0de51d`,
SHA-256 `bc42798f059416957a4e67cefea3b469cb2c08caa1fc2d73997f81e76a362fe1`).
`0.1.4` remains tag `0.1.4` @ `9a07e9a33726dd187080b00a44505f9bbd31bd97`,
SHA-256 `15c4220df078de9c350aea98e0f9187890cec066e4d3ee242170a2ab762ed80f`.
macOS 12 Monterey Intel is deliberately unsupported.
## 0.1.6 release notes
Reopen keeps the restored index cache when a DiskANN sidecar file belongs to
an older generation. A failed sidecar rewrite leaves that previous file in
place; it no longer discards the other cached indexes. A corrupt or truncated
sidecar for the current generation still misses and rebuilds.
## 0.1.6 post-publish checklist
1. Hosted CI run `35754487571` on revision `c7b828c` is green.
2. The published crate SHA-256 matches
`67c238a010add5d35d085a87a9183b9d3e08aef6a89b1b5edf19186b31b65062`.
3. Formal git tag `0.1.6` points at that revision, and `cargo publish`
uploaded the matching crate.
## 0.1.5 historical binding
`a3s-vec` `0.1.5` is published. Tag `0.1.5`, and crates.io SHA-256
`bc42798f059416957a4e67cefea3b469cb2c08caa1fc2d73997f81e76a362fe1` bind to
revision `84f8985a669d0e4d33eb0c4884c7ccb90e0de51d`. Hosted CI run
`35746396350` is green for that revision. The prior `0.1.4` package remains
on the registry as a historical artifact (tag `0.1.4` @
`9a07e9a33726dd187080b00a44505f9bbd31bd97`, SHA-256
`15c4220df078de9c350aea98e0f9187890cec066e4d3ee242170a2ab762ed80f`). macOS 12
Monterey Intel is deliberately unsupported.
## 0.1.5 release notes
This patch release keeps the public query contract (`f64` scores, exact
re-rank, HNSW `ef` 64, `Durability::Always`, no default IVF `scale_factor`)
and removes physical work that tests can still fail:
- An `Always` commit is acknowledged only after its durability sync, and that
sync does not hold the published-state lock.
- A failed DiskANN sidecar write still leaves the other derived indexes
restorable from the cache when the sidecar file is absent. A leftover file
from an older generation is corrected in `0.1.6`. A corrupt sidecar still
misses and rebuilds.
- A checkpoint of changed documents writes a format-5 delta of the changed
bodies and keeps the base snapshot. Published format-4 snapshots still open.
A second flush of an already checkpointed revision writes nothing.
- HNSW neighbor scoring may run in parallel and still publishes the ordered
single-thread `f64` neighbor sets.
## 0.1.5 post-publish checklist
1. Hosted CI run `35746396350` on revision `84f8985` is green.
2. The published crate SHA-256 matches
`bc42798f059416957a4e67cefea3b469cb2c08caa1fc2d73997f81e76a362fe1`.
3. Formal git tag `0.1.5` points at that revision, and `cargo publish`
uploaded the matching crate.
## 0.1.4 historical binding
`a3s-vec` `0.1.4` is published. Tag `0.1.4`, and crates.io SHA-256
`15c4220df078de9c350aea98e0f9187890cec066e4d3ee242170a2ab762ed80f` bind to
revision `9a07e9a33726dd187080b00a44505f9bbd31bd97`. Hosted CI run
`35510190796` closed that gate. The prior `0.1.3` package remains
on the registry as a historical artifact (tag `0.1.3` @
`88599126a4c179d8a0df24bd52963d372ea8eb67`, SHA-256
`c5c692f409c4870048a5f081f66f9072893af2047042487835f97b1d6aa3d9f2`).
## 0.1.4 release notes
This patch release keeps the public storage and query contracts and turns
persistence DoS ceilings into first-class configuration:
- New public type `StorageCeilings` with product defaults **8 GiB** snapshot /
index-cache / WAL-replay and **512 MiB** DiskANN sidecar (same magnitudes as
`0.1.3`).
- Process default via `ConfigBuilder::storage_ceilings`; per-collection
override via `CollectionOptions::set_storage_ceilings`. Zero is rejected;
omitted fields keep the product default.
- Encode, open, restore, and DiskANN attach/validate paths honor the captured
ceilings end-to-end (raised limits no longer break sidecar restore).
- Protocol/format guards stay hardcoded elsewhere (1 MiB manifest, 64 MiB
single WAL frame, 4 KiB lock owner).
- The engine still never autodetermines ceilings from host RAM or free disk.
Do not lower `ef`, drop exact re-ranking, or switch public scores to `f32` to
manufacture a benchmark win.
Supported platforms: Linux x86_64/aarch64, Windows x86_64, and macOS
arm64/x86_64 on current hosted images (macOS deployment target 15.0).
## 0.1.3 release notes (historical)
This patch release kept the public storage and query contracts and raised the
finite storage DoS ceilings so million-document dense corpora can flush and
persist derived indexes on workstation hosts:
- Document snapshot write/recovery ceiling: **512 MiB → 8 GiB**.
- Derived index cache payload and on-disk index-cache file ceiling: **512 MiB →
8 GiB** (aligned with snapshots so HNSW graphs can persist beside the
authoritative corpus).
- Committed WAL replay ceiling: **512 MiB → 8 GiB**.
- DiskANN sidecar file ceiling remains **512 MiB** (unchanged).
- Honest same-host million-scale directional evidence under the existing
fairness harness (`RAYON_NUM_THREADS=1`, protocol knobs unchanged): see
[README.md](README.md) and [BENCHMARKS.md](BENCHMARKS.md). Protocol defaults
are not tuned for high million-scale recall; do not market recall@10 from
that table as an accuracy SLO.
## 0.1.2 release notes (historical)
- Character-trigram prefilter for FTS wildcard/fuzzy matcher expansion.
- Packed dense Flat acceleration; parallel Flat Cosine under the default Rayon
pool; bit-identical `f64` public scores and exact re-ranking retained.
- First-principles a3s-vec ↔ zvec remeasure protocol and Apple Silicon medians.
## Public API review
The release-facing contract has the following boundaries:
- Filter parsing, tokenization, and index quantization are implemented in this
crate and cannot be named through the public crate surface.
- Collection and process configuration use typed Rust values. Unsupported
controls and index/query combinations fail with typed errors before
mutation.
- Persistence DoS ceilings use `StorageCeilings`; document/query resource
budgets remain separate as `CollectionResourceLimits`.
- Schema backfills and candidate-schema validation accept a bounded typed
worker count through `AddColumnOption`/`AlterColumnOption`; the effective pool
is capped by work size, host parallelism, and 256 workers, while publication
stays atomic and deterministic across worker counts.
- Public embedding and query-executor ports require `Send + Sync`. The
`public_api_contract` integration test also enforces `Send + Sync` for the
owned public handles, schemas, queries, values, statistics, errors, and
`StorageCeilings`.
- `unsafe_code = "deny"` remains active. The mmap option is an immutable
anonymous snapshot of a fully validated sidecar, not a mutable file-backed
mapping.
- `version()`, the numeric version accessors, and `check_version()` are checked
against the package's `0.1.8` identity.
- The public feature matrix checks every advertised query/lifecycle route,
all six ANN families across their supported metrics (including metric-aware
Vamana and DiskANN/PQ), cache/sidecar reopen, and the explicit binary-query
boundary against deterministic fixtures. Binary32/Binary64 radius and
projection/include-doc-id combinations also have asserted matrix rows. Its
smoke-scale feature-matrix, concurrent-reader, mixed-workload,
scale-comparison, and lifecycle-matrix performance CSVs are required hosted
CI artifacts; same-host p50/p95/p99 baselines are
recorded in [`BENCHMARKS.md`](BENCHMARKS.md).
- The locked dependency graph passes `cargo audit --deny unsound`: no known
vulnerability or unsoundness advisory is present at the candidate revision.
The audit still reports four upstream maintenance warnings (`bincode`,
`bitmaps`, `fxhash`, and `paste`); they are non-blocking because no patched
release or unsound finding is currently available for those paths, and the
exact versions remain pinned in `Cargo.lock`.
- `SearchQueryBuilder` dense and pure-FTS routes are executed against the same
collection oracles, and `include_doc_id` is checked for deterministic,
generation-local ordinals across projections and reopen.
## Reproducible candidate artifact
After every required hosted CI job passes on `main`, the `Versioned release
candidate` job runs `cargo package --locked`. It uploads these files in one
revision-bound Actions artifact:
- `a3s-vec-0.1.8.crate`;
- `a3s-vec-0.1.8.crate.sha256`;
- `a3s-vec-0.1.8.release.json`, which records the package version, source
revision, workflow run, and build runner.
- `feature-matrix.csv`, `concurrent-queries.csv`, `mixed-workload.csv`,
`scale-compare.csv`, and `lifecycle-matrix.csv`, which record the
smoke-scale correctness/performance gates, including management-plane
lifecycle, resource, and maintenance operations.
- One `a3s-vec-platform-performance-<platform>-<revision>` directory for each
hosted platform, containing the same five validated smoke CSVs. These
artifacts show whether the metrics and recall gate hold across the supported
OS/architecture matrix.
The same package can be reproduced locally without changing external state:
```text
cargo fmt --all -- --check
cargo clippy --locked --all-targets --all-features -- -D warnings
cargo audit --deny unsound
cargo test --locked --all-features
cargo test --locked --test feature_matrix
cargo bench --locked --bench feature_matrix --features async
A3S_VEC_BENCH_SCALE=smoke cargo bench --locked --bench feature_matrix --features async
cargo bench --locked --bench concurrent_queries
A3S_VEC_BENCH_SCALE=smoke cargo bench --locked --bench concurrent_queries
cargo bench --locked --bench mixed_workload
A3S_VEC_BENCH_SCALE=smoke cargo bench --locked --bench mixed_workload
cargo bench --locked --bench scale_compare
A3S_VEC_BENCH_SCALE=smoke cargo bench --locked --bench scale_compare
cargo bench --locked --bench lifecycle_matrix
A3S_VEC_BENCH_SCALE=smoke cargo bench --locked --bench lifecycle_matrix
cargo doc --locked --no-deps --all-features
cargo package --locked --allow-dirty
```
## 0.1.4 post-publish checklist
1. Hosted CI run `35510190796` on revision `9a07e9a` (binding confirmed when
green).
2. The published crate SHA-256 matches
`15c4220df078de9c350aea98e0f9187890cec066e4d3ee242170a2ab762ed80f`.
3. Formal git tag `0.1.4` points at that revision, and `cargo publish`
uploaded the matching crate.
## Deliberate non-support
- macOS 12 Monterey on Intel x86-64 is unsupported. The former
`macOS 12 Intel Runtime Qualification` workflow and host-fenced script have
been removed. A `MACOSX_DEPLOYMENT_TARGET=12.0` build is not a supported
configuration.
- Native async file reads and direct file-backed mmap remain refused until an
invariant test fails (see VEC-R2 in [`ROADMAP.md`](ROADMAP.md)).