test: hnsw recall@10 test flakes on CI (0.58-0.65 vs ~0.98) — parallel_insert makes the graph nondeterministic #302

hnsw_index::tests::test_hnsw_recall_at_10_on_3000_random_vectors (src/hnsw_index_tests.rs:82) sometimes fails on CI with recall far below its 0.95 bar. Each run's recall varies even though the data seeds are fixed.

Evidence

The data uses seed 42 (3000 rows, dim 32) and the queries use seed 4242 (100 queries). m=16, ef_construction=200, ef_search=64, cosine distance.

CI failures:

• 0.654: run 37276802109, job 111660276084, branch fm/fix-parser-depth (fix(parser): reject over-deep and oversized input instead of overflowing the stack #295). The same commit passed this test in two earlier attempts of that run.
• 0.577: run 37234114883, branch fm/en-en8.

Local runs (debug, dev box), 20 runs isolated and 20 runs with 5 copies at once:

Likely cause

HnswIndex::build inserts with hnsw_rs parallel_insert (src/hnsw_index.rs:107). That makes the graph depend on thread scheduling, so recall varies even with fixed data. On a loaded CI runner, an occasional build produces a much worse graph. Usual recall is about 0.98; the bad builds dropped to 0.58-0.65.

Options

• Build the test index deterministically (serial insert) and keep the 0.95 bar.
• Find out why some parallel builds come out degenerate. A 0.58 recall at ef_search=64 may point to a real build-quality problem under contention, not just test noise.

The test was not weakened in PR 295.