turbovec-sqlite sqlite-v0.3.0

0.3.0 — 2026-10-11

• Commit inserts and deletes without loading the compressed index. A new
  <name>_ids shadow table answers rowid conflict checks, rowid lookups, full
  rowid scans, and count(*); a new <name>_info row holds the row count and
  delta counts a commit needs. Connecting reads that one row. The index loads on
  the first search, or when a commit compacts. On a 1,000,000-row,
  1,536-dimensional index (745 MiB base), a one-row delete-and-insert commit
  from a fresh process fell from 468 ms and 830 MiB peak RSS to 0.8 ms and
  24.5 MiB, with a warm page cache. The first search costs what it did.
• Upgrade 0.2.0 storage in place on its first write: the commit creates _info
  and _ids from the index it already loaded and empties <name>_meta. This
  took 1.1 s on the index above. Reads alone do not upgrade. The upgrade is
  one-way: 0.2.0 then refuses the index with "Query returned no rows", including
  a connection opened before the upgrade, so it cannot write without
  maintaining _ids. Chunk and delta formats are unchanged.
• Record a storage format number in _info and refuse an unknown one by
  number.
• Fix a savepoint rollback that lost a replaced row. On a base under 1 MiB, an
  insert after an INSERT OR REPLACE dropped its vector, so a later delete took
  its destructive checkpoint after the replacement, and rolling back past the
  checkpoint undid the replacement by removing the row. Inserts now keep their
  vectors after any delete or replacement. 0.2.0 has this bug. The test script now
  runs the transaction model check for 300 seeds of 30 steps, which reproduces
  it; the previous 20 seeds of 20 did not.
• Pin the extension on Windows so it stays loaded after its last connection
  closes. SQLite unloads it then, but every search starts TurboVec's rayon
  worker threads, which keep running code from the library and fault once
  Windows unmaps it. Linux and macOS already keep it loaded. 0.2.0 has this
  too; a test now checks that the library outlives its connections.
• PRAGMA integrity_check compares _ids and the _info counts with the
  index. A load checks the index's row count against _info.
• Each insert and delete writes one _ids row, which total_changes() counts.
  A 100,000-row bulk insert in one transaction took 11% longer (1.29 s to
  1.43 s).