Flaky test: TestFindNearestApproximate first result can be y instead of x (PQ on three vectors) #149

tests/test_vector_index_introspection.py::TestFindNearestApproximate::test_returns_the_nearest_record_and_its_distance failed once on PR #148 (job https://github.com/ArcadeData/arcadedb-python/actions/runs/40999879904/job/123882125988, Test arcadedb (linux/amd64, Python 3.14)): assert record.get("name") == "x" got "y". The same test passed on every other leg of that run and on the last six runs of test-jvm.yml on main (scheduled run of 2026-10-30 03:04 UTC included), and PR #148 does not touch the PRODUCT path, so this looks like a flaky test rather than a regression.

Why it can flip: the fixture builds a PRODUCT index on three 3-dimensional vectors (PQ_VECTORS) with pq_clusters=4 and asks find_nearest_approximate([0.9, 0.1, 0.0], k=1). The codebook is trained on those few vectors, and the approximate answer is ranked on the PQ codes, so whether x or y comes first can depend on how the training lands. test_k_results_come_back_nearest_first asserts the same first name with k=3.

Possible fixes (to choose in the PR): rerank so the returned distance is exact (rerank_oversample, or a k large enough that the exact rerank covers every row), vectors far enough apart that any codebook of 4 clusters separates them, or an assertion on the set of the top k rather than the first name. A fix should be checked by repeating the test many times on the same leg, not by one green run.