Apply distance constraints in rescore KNN queries (fixes #308)

[stumpylog]

Summary

Fixes #308. On vec0 tables created with INDEXED BY rescore(...), KNN queries silently ignored distance <op> ? constraints:

SELECT rowid, distance FROM t
WHERE embedding MATCH :vector AND k = :limit AND distance <= :threshold
ORDER BY distance

This returned the top-k by distance without filtering out rows above the threshold.

Root cause

The standard chunk-scan path filters candidates against the distance constraints parsed from idxStr/argv. Rescore columns dispatch to rescore_knn, which received those constraints but never applied them.

Fix

Apply the constraints in rescore_knn phase 2, after the exact float distances are computed and before top-k selection — so they target the final rescored distance (what the distance column reports), not the coarse quantized distance from phase 1. Candidates failing a GE/GT/LE/LT constraint are dropped, with an early-out when none survive.

Testing

• Added test_knn_distance_constraint_le and test_knn_distance_constraint_lt_gt to tests/test-rescore.py; both fail on main and pass with this change.
• Full loadable suite: 484 passed, 132 skipped.

🤖 Generated with Claude Code