perf: float vectors in a plain table, allocation-free exact scan

[Verdenroz]

Description

sqlite-vec reads the whole 4 MB vector block for a point lookup, and for a rowid IN set on a KNN query as well, so the two-stage path could never beat exact search by much: 54ms against 68ms at 22k chunks. Floats now live in a plain rowid table and the vec0 float table is gone. The coarse pass rescores 400 candidates in 29ms at recall@10 of 1.000, exact search is a Rust scan that measures 12.0ms at 16k chunks in the soothfast bench, and the file is 151 MB after VACUUM instead of 157.

Changes

• Added chunk_vectors with ON DELETE CASCADE and a v2 to v3 migration that copies vectors out of vec_chunks and drops it. The copy takes 5.7s on 22k chunks and the file sits at 265 MB until the next VACUUM.
• Replaced the vec0 exact path with a scan that computes cosine on the blob in place with a bounded top-k, so a full scan allocates nothing per row. The candidate id list for the rescore is built in one allocation.
• Added repos.chunk_count, maintained by triggers on chunks, in place of the per-query count(*) join.
• Refused files whose schema_version is newer than the binary, since v3 drops a table an older binary would still query.
• Made refiner_curve recall distance-based and added recall@10. 1,510 of 22,581 live vectors are exact duplicates, and id-based recall read tie reordering as a 4% loss.

Testing

• cargo test --workspace: 57 passed, 0 failed. cargo clippy --workspace --all-targets -- -D warnings clean. The migration test now drives a v1 file through to v3 and checks the copied vectors and chunk count.
• Refiner curve on a fresh copy of the live index, cross-repo: exact 70ms; coarse 100/200/400/800/1600 at 9/16/29/63/75ms with recall@10 of 0.998/1.000/1.000/1.000/1.000 and recall@50 of 0.918/0.980/0.994/0.998/1.000.
• Eval on a same-moment copy, HyDE off: identical to the base branch on all three sets.
• cargo soothfast measure -p scry-core --save-baseline base on a quiet machine, and an exact docs/search.md claim that a dense query over 16k chunks allocates at most 11 times. cargo soothfast docs check passes.
• Deploy order matters: install the new binary and restart the service together. An old binary opening a migrated file fails at query time (vec_chunks is gone); the new binary refuses files newer than itself.