[MayCXC]

A KNN query on an IVF column ignores a rowid in (...) filter and returns the unfiltered k nearest rows.

create virtual table v using vec0(a float[8] indexed by ivf(nlist=4, nprobe=1));
-- 40 inserts, then compute-centroids; :q is row 3's vector
select rowid, distance from v where a match :q and k = 3 and rowid in (1, 2);
-- 3 (0.0), 30 (1.379), 10 (1.4); a flat column returns 1 (2.099), 2 (2.929)

vec0Filter_knn() collects the rowids the filter names, and vec0BestIndex() tells SQLite to omit the constraint, but the IVF dispatch never passed them to ivf_query_knn(). It now does, and a filter's rows are ranked by their exact distances, read through ivf_get_vector_data(), skipping a rowid the table does not have. The cell scan and the oversample re-rank run only without a filter: the probed cells hold the rows near the query, which the filter may not name, so a filter applied to them alone would miss the named rows in other cells (above, row 2 is in a cell that nprobe=1 does not probe).

This builds on #349 and carries its commit, since on main ivf_get_vector_data() returns a quantized column's quantized bytes rather than its vectors.

Test: test_ivf_knn_rowid_in in test-ivf.py compares rowids and distances with a flat column built from the same vectors, before and after compute-centroids, for filters that name rows outside the probed cell and a rowid the table does not have, parametrized over no quantizer, int8 and binary. All three fail on main and on #349 and pass here. The Python suite passes (with -DSQLITE_VEC_EXPERIMENTAL_IVF_ENABLE=1, 306 passed, 4 skipped; without it, 239 passed, 71 skipped), under ASan and UBSan as well, and both builds' warnings are main's.

The distance half is #316, which applies distance constraints in every backend. The two merge cleanly, and together a distance constraint applies to the k rows the filter ranked, as it does to the probed cells' results, so rowid in (...) and distance > x can return fewer than k of the named rows where a flat column returns k. This also merges cleanly with every open PR that touches the IVF code, #313, #343 and #347 included.

Origin. Found in the IVF port to photostructure/sqlite-vec, photostructure#12 (65817f4), whose ivf_knn() takes the rowid filter and distance constraints of the rescore index's KNN; the change and its test are written here for asg017's main. No open or closed PR or issue here makes the same change.