Linear-time inserts, Turso v0.8.0-pre.13 vector parity, browser vector gate

[Marc-André Moreau (mamoreau-devolutions)]

Summary

Vector search parity. Against Turso's Rust engine, Ahtola's vector support was already complete: the vector functions, the serialized formats, and an exact USING vector index for ORDER BY vector_distance_*(…) LIMIT k. This PR moves the turso-src pin to v0.8.0-pre.13 (64b8ef574) and ports the one behavioral change in that release:

• Dense float32/float64 vector_distance_cos follows b9414023d: two zero vectors are at distance 0, and a zero dot product is at distance 1, decided before dividing. This is Turso's pure-Rust fallback, which Turso runs on wasm, so native and browser results stay identical.
• The rest of the core/vector delta is simsimd build plumbing. turso/vector.sqltest is byte-identical to pre.7.
• Upstream's index-method OFFSET fix (78395e704, #7523) needed no engine change here. Its FTS case is added as a test.
• The libSQL-only features from Turso's vector-search blog post (F32_BLOB(N) typing, libsql_vector_idx, vector_top_k, DiskANN) do not exist in the Rust engine through pre.13, so they are out of scope.

Insert performance. Every statement did O(table) work, so row-at-a-time inserts were quadratic:

• Catalog clones deep-copied every row of every table on every statement. Row and rowid lists are now shared copy-on-write in 1,024-element chunks (CowChunkedList).
• Two sites patched stored rows in place: incremental blob writes and sqlite_sequence renames. They now replace the row instead, because stored row arrays are shared with clones.
• PrepareInsert and ValidateRowids rebuilt a HashSet of every rowid per statement. Rowids are now cached per table as an immutable set plus their maximum, keyed on row-store lineage, revision and count, and extended by inserts.
• INSERT validation checks NOT NULL and non-unique index keys only on the new rows, as SQLite does. It skips rescanning a rowid alias that ValidateRowids just proved unique, and validates through a view instead of copying the table. Unique constraints are still checked against every row.

Workload	Before	After
64-dim vectors with an index, one row per INSERT, 20k rows	9,120 µs/row	117 µs/row (87 at 40k)
Integer rows in one transaction, 20k rows	2,236 µs/row	27 µs/row (12 at 40k)

One visible behavior change from the cosine port: float32 vectors whose squared norm underflows to zero, while their dot product doesn't, now report +∞ instead of 1. ManagedVectorIndexReviewRegressionTests.CosineRecallSurvivesScalarNormCollapse is updated for that. It still requires the index and the scan to agree, and now also compares at full-table LIMIT.

Browser. samples/BrowserWasmConsumer now builds a USING vector index over OPFS, reopens it, and requires the planner to choose it and to return the same rows as the unindexed scan. The expected status in scripts/Invoke-BrowserPackageConsumer.ps1 is extended with vector-search=195-1987-643.

Testing

• Full managed suite on net10.0: 8,246 passed, 0 failed. Core builds with zero warnings on net8.0, net9.0 and net10.0, with the AOT/trim analyzers on.
• New tests: CowChunkedListTests (randomized differential against List<T> with clones), RowStoreCopyOnWriteTests (clone isolation, failed-statement atomicity, AUTOINCREMENT rename rollback, blob write rolled back to a savepoint), cosine zero-norm and zero-dot cases, and FTS ranked LIMIT/OFFSET.
• The browser consumer app was run in Chromium through project references and produced the exact expected PASS string. The Playwright gate itself was not run locally.
• validate-project-closure passes. format-check reports only pre-existing issues in lines this PR doesn't touch.

🤖 Generated with Claude Code