turbovec-sqlite sqlite-v0.2.0

0.2.0 — 2026-09-26

• Bring a warm reader forward by replaying only the delta batches other
  connections committed, instead of rebuilding the whole index. A reader's first
  query after a concurrent one-row commit dropped from 7.6–12.3 ms to about
  1.4 ms on a 20,000-row by 1,536-dimension index; the saving grows with index
  size. A compaction, a gap in the delta chain, or any other surprise still
  falls back to the full read.
• Cap the delta budget at the base image. A high-dimensional index pays a fixed
  rotation and codebook cost — an empty 1,536-dimension base already serializes
  to 1.67 MiB — so the 16 MiB floor alone let small tables accumulate raw
  float32 deltas several times larger than the 4-bit base they described. A
  fresh 2,000-row by 1,536-dimension table now stores 3.21 MiB instead of
  13.97 MiB, and its cold open plus first query fell from 27 ms to 1 ms. Small
  commits on a large index are unchanged.
• Stream base chunks straight into TurboVec's loader instead of materializing
  the image first. IdMapIndex::from_bytes copies what it is handed, so opening
  an index held two full copies; peak growth on a 48.5 MiB image fell from
  104 MiB to 61 MiB.
• Plan joins whose MATCH is unusable in one loop order. xBestIndex now
  declines those orders instead of failing the statement, so batch KNN over a
  table of queries works: queries q join v on v.embedding match q.embedding and v.k=n. A query no order can serve reports SQLite's "no query solution"
  rather than a tailored message.
• Report the planned top-k as the KNN plan's estimated row count instead of a
  fixed 10, and keep xBestIndex off the blocking path when another statement
  holds the warm index.
• Free the previous virtual-table error message before replacing it. SQLite's
  savepoint callbacks never reclaim it, so a failing xSavepoint, xRelease,
  or xRollbackTo leaked its message.
• Fail the build if a future libsqlite3-sys grows sqlite3_module, which
  would silently stop PRAGMA integrity_check from reaching turbovec0.
• Assert in the transaction model check that each rowid holds the vector version
  the model expects, not only that the rowid exists. Exercise turbovec0
  creation, insertion, search, and damaged-storage reporting from the
  statically linked C smoke test.
• Say why turbovec0 is unavailable on a host below SQLite 3.38, naming the
  required and found versions, instead of reporting a missing internal callback.
  A statically linked build inherits whatever libsqlite3 it was linked
  against, so this is the message that path produces; the scalar turbovec_*
  functions keep working there.