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 right on the bar and 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

The test data is already deterministic (generate_random_array_with_range seeds StdRng::from_seed([13; 32])), but the index build is not. RaBitQ draws a random rotation at build time from an unseeded entropy RNG (random_fast_rotation_signs / random_orthogonal), and for nlist=4 the IVF k-means init is unseeded as well. One rotation is drawn per build and shared by every query, so recall varies run to run around the 0.5 bar. With 1-bit RaBitQ over uniform-random 32-dim data the mean recall sits at ~0.5, so the assertion is inherently borderline (the test already carried a "RQ doesn't perform well for random data" note and a commented-out #[ignore]).

Fix

Keep the strict >= 0.5 assertions and remove the flakiness by making the build deterministic through existing build-param seams, with no production change. The rotation is pinned via the existing RQBuildParams.rotation field (the same seam distributed builds use to share a rotation): the test constructs a fixed RabitQuantizationMetadata with fixed sign bits for the Fast rotation and a deterministic Haar-like orthogonal matrix (seeded Gaussian + Gram-Schmidt) for the Matrix rotation. The IVF centroids are pinned via IvfBuildParams::try_with_centroids, which bypasses k-means and removes the nlist=4 variance.

The change is confined to #[cfg(test)] mod tests. A no-op guard (determinize_rq_params returns params unchanged unless the last stage is RQ) leaves the SQ/PQ/flat/HNSW tests that share the same helpers unaffected.

Verification

All 12 test_build_ivf_rq permutations pass, including the previously-failing case_2::rotation_type_1_RQRotationType__Fast. Ran the matrix three times back-to-back with identical results. test_build_ivf_sq and test_build_ivf_pq still pass (shared-helper no-op path). cargo fmt --all and cargo clippy -p lance --tests -- -D warnings are clean.