v0.4.2

Added

• nitrite_vector — new approximate-nearest-neighbour (ANN) vector index and RAG store
  extension crate, with two backends (in-memory HNSW and disk-resident DiskANN) selectable per
  database or per index, cosine/Euclidean/dot metrics, and F32/F16/I8 stored-vector
  precision.
• nitrite_vector — automatic index rebuild. Both backends detect stale/corrupt index
  storage on open (torn writes, missing/corrupt sidecar, crash before checkpoint, format
  change) and transparently re-index from the collection's documents.
• nitrite_vector — per-index configuration. VectorModule::builder(...).index_config( collection, field, config) lets collections with different embedding dimensions, metrics,
  precisions, or backends coexist in one database.
• nitrite_vector — DiskANN crash-consistency machinery: data-file dirty bit + generation
  stamping, checksummed sidecar replaced atomically (tmp + rename), structural validation of
  everything read from disk, and a periodic sidecar checkpoint (every ~8k mutations) in
  addition to flush-on-close.
• nitrite_vector — HNSW deletes now reconnect the orphaned neighborhood (diversity-pruned),
  so sustained insert/delete churn no longer fragments the graph.

Fixed

• nitrite — range filters mishandled null values. Value's mixed-type ordering falls
  back to string comparison, which made Null compare as the string "null". As a result:
  • collection-scan gt/gte (and, depending on the store's key order, other range filters)
    matched documents whose field is null or missing;
  • indexed lt/lte returned an empty result as soon as the indexed field was null in
    any document, because the forward scan seeded from the null index key and terminated
    immediately.
    Range filters now explicitly treat null/missing values as never lesser or greater than the
    search term: the document-scan path rejects null field values, and every index range scan
    (all four comparison modes, both scan directions, plus the bounded ceiling(lower) .. floor(upper) scan) skips the null index key. Mirrors nitrite-java issue
    #1262.
• nitrite-int-test — added regression tests for range scans over indexed, unique-indexed
  and non-indexed fields containing nulls (index_null_key_scan_test.rs).
• nitrite_vector — torn HNSW persists could panic searches. Each mutation now persists as
  a single atomic put_all batch (records + header), traversal skips dangling neighbor
  references instead of panicking, and loading sanitizes the graph (bad records dropped,
  dangling links pruned, entry point re-elected).
• nitrite_vector — DiskANN writer races. Inserts, removes, delete-consolidation, and the
  PQ encode step now serialize on a per-index write gate (background passes take it per chunk),
  eliminating lost adjacency updates and stale-edge reattachment; only slots pending at the
  start of a consolidation sweep are reclaimed.
• nitrite_vector — PQ training runs in the background (single-flight) instead of stalling
  the insert that crosses the threshold; queries fall back to exact distances for nodes not yet
  encoded, so they can never become invisible mid-training.
• nitrite_vector — DiskAnnIndex::flush no longer permanently disables background
  consolidation; it is now a safe, repeatable checkpoint.
• nitrite_vector — the HNSW backend now honors the configured stored-vector Precision
  (F32/F16/I8) as documented; adjacency-only changes no longer rewrite full vectors, cutting
  per-insert write amplification by roughly an order of magnitude; persistence happens outside
  the graph lock so searches are never blocked on storage I/O.
• nitrite_vector — a racing double-open of the same index could create two live instances
  over the same storage (double-checked registry lock); DiskANN file names are sanitized +
  digest-suffixed so hostile collection/field names cannot escape the database directory;
  min_score now over-fetches (4×k) before applying the cutoff so it no longer starves k.