ParqDB 0.2.0rc3

0.2.0rc2 was rejected by installed-wheel verification before PyPI publication. 0.2.0rc3 contains that feature set plus the release-gate and dependency fixes below.

[0.2.0rc3] - 2026-08-21

Fixed

• The vendored Arrow Parquet runtime now uses Apache Thrift 0.23, removing the
  vulnerable unbounded compact-protocol varint decoder from release wheels.
• Installed-wheel tests now resolve repository fixtures independently of the
  process working directory, preserving source-tree isolation during release
  verification.
• Pull requests now build and exercise an installed wheel on Python 3.11,
  including the embedded DataFusion namespace-coexistence check, before a
  release tag can be created.

[0.2.0rc2] - 2026-08-21

Added

• LVQ4/LVQ8 builds now emit a strict, immutable top-level manifest.json and
  package-local centroid objects, so one snapshot directory can be copied
  directly to public HTTP storage without repacking.
• The new @parqdb/browser client performs global leaf-centroid routing,
  explicit postings row-group selection through HTTP Range requests, and
  LVQ distance/top-k execution in WebAssembly, returning source keys plus
  _distance without listing objects or joining the source table.
• A new parqdb publish command builds or reuses a static IVF-LVQ4/LVQ8
  index, publishes its source table and model assets to an immutable local or
  S3-compatible prefix, and verifies public HTTP Range and CORS behavior.

Changed

• IVF training now defaults to a persisted two-level hierarchy. Postings are
  grouped into root-aligned cid_bucket files, discovered through an immutable
  manifest, and pruned with explicit Parquet row-group access plans. Root
  training samples 512 points per root, re-seeds empty roots from populated
  donor partitions, then assigns the exact leaf budget proportionally to the
  sampled root populations. It falls back after bounded empty-root recovery to
  flat leaf training followed by a synthetic valid hierarchy.
• Python index discovery and removal now stay behind the table-centered API;
  internal catalog and native repository bridges are no longer exposed through
  session facade attributes.
• Browser queries coalesce and parallelize HTTP Range reads while keeping
  metadata prefetch and the in-browser byte cache bounded.
• The optional publication stack uses the smaller tokenizers runtime instead
  of the full Python Transformers package while preserving pinned MiniLM token
  and embedding parity.

Fixed

• Publication build work is now bound to the exact source and configuration,
  so a changed input cannot silently reuse a stale index. Interrupted embedding
  output is committed atomically and cannot be mistaken for a complete source.
• Static publication cleans partial local and S3 writes on failure and reports
  post-commit public HTTP verification failures without suggesting an unsafe
  retry against the same immutable destination.
• Dependency auditing now covers every locked optional extra in CI and release
  jobs; the vulnerable h2 and Transformers dependency versions are removed.

Removed

• The pre-1.0 Session.to_arrow() compatibility alias; use collect() for a
  materialized Arrow table or stream() for incremental batches.