What

test_build_ivf_rq (rust/lance/src/index/vector/ivf/v2.rs) is flaky: it builds an IVF_RQ index over seeded random vectors and asserts recall >= 0.5, but recall sits close enough to the bar that it occasionally dips below. It surfaced in CI on #7371 as recall: 0.49 for the case_2::rotation_type_1_RQRotationType__Fast permutation (nlist=1, Cosine, Fast rotation) - see the failed linux-build job: https://github.com/lance-format/lance/actions/runs/28875488220/job/85649319644. It is a pre-existing flake, not caused by any recent change to the search path.

Root cause

test_recall issues its queries with nprobes == nlist, so every partition is scanned and the measured recall reflects RaBitQ quantization error alone. The test builds with num_bits = 1, which means ex_bits = 0: only the sign of each rotated coordinate survives. Over 512 uniformly random, L2-normalized 32-dim vectors the top-100 neighbours are barely distinguishable at that resolution, so recall lands near the bar rather than far above it.

Measured across 336 samples (3 runs x 6 rstest cases x both rotation types x the f32/f64/remap/multivec recall sites):

At num_bits = 1 the 0.5 bar sits at mean - 4.1 * stdev. The dataset itself is already deterministic (generate_random_array_with_range seeds StdRng::from_seed([13; 32])), but the build draws a fresh random rotation every run from an unseeded entropy RNG, and the IVF k-means init is unseeded too, so that tail is reachable. Seeding those would only fix one draw of a distribution whose mean is 0.667; it would not make the assertion mean anything.

For comparison, the sibling tests in the same file assert IVF_FLAT 1.0, IVF_PQ 0.9 / 0.9 / 0.85, IVF_SQ 0.85 / 0.85 / 0.75.

Fix

Build with num_bits = 5 instead of 1 and raise the assertion from 0.5 to 0.9, which lands at mean - 6.9 * stdev. No production change; the diff is confined to #[cfg(test)] mod tests.

ex_bits = 4 also exercises a FastScan ex-code kernel that test_build_ivf_rq_multi_bit_persists_split_codes_and_searches never reaches: its num_bits 4 and 6 cases have ex_bits 3 and 5, which take the bit-plane repack path, and its num_bits = 9 case has ex_bits = 8. supports_ex_fastscan accepts 2 | 4 | 8.

Both rotation types stay in the matrix. The commented-out #[ignore = "Temporarily skipping flaky 4-bit IVF_RQ tests"] line is removed: it was dead, and it referred to 4-bit tests when the test was 1-bit.

Verification

All 12 test_build_ivf_rq permutations pass over 10 consecutive runs (120 invocations), including the previously-failing case_2::rotation_type_1_RQRotationType__Fast. The rest of index::vector::ivf::v2::tests passes, which covers the shared test_index / test_remap / test_index_multivec helpers used by the SQ, PQ, flat and HNSW tests. cargo fmt --all is clean in all three workspaces and cargo clippy -p lance --tests -- -D warnings is clean.