flaky: ruvector-quantization 'should perform search with binary quantization' recall floor fails on unseeded random vectors

Start with v3/@claude-flow/plugins/tests/ruvector-quantization.test.ts, especially the binary-quantization test at line 493 and the related perturbation case near line 508. Run npx vitest run __tests__/ruvector-quantization.test.ts, then make the recall check deterministic or base it on the planted near-duplicate scenario described in the issue. Done means the test passes reliably without masking regressions or depending on unseeded random vectors.

Observed: ruflo CI run 33974289751 (PR #3184 at 8b3927d, ubuntu Test Suite) — the CI test ratchet flagged one unexpected failing file, v3/@claude-flow/plugins/__tests__/ruvector-quantization.test.ts:

RuVector Quantization > Binary Quantization > should perform search with binary quantization
AssertionError: expected 0 to be greater than or equal to 0.1
  at __tests__/ruvector-quantization.test.ts:493:22

The PR does not touch that file (diff vs main: LEDGER, gist, cli-cold-start test/bench, ADR status summary). The file is byte-identical to main, and locally it passes 28/28 (npx vitest run __tests__/ruvector-quantization.test.ts). It also passed on #3204's run 33972961703 minutes earlier on the same base.

Why it flakes: the test builds its corpus with randomVector(dimensions) (unseeded Math.random) and asserts binary-quantized top-k recall ≥ 0.1 against exact search. Binary quantization over i.i.d. random vectors has no such floor for small k — recall of 0 is a legitimate outcome of the RNG, not a regression. Sibling tests in the file have the same shape (line 508 perturbs with Math.random() too).

Ask (either):

1. seed the generator (a small LCG/xorshift randomVector(dim, seed)) so the corpus is fixed and the assertion is deterministic, or
2. assert a property binary quantization actually guarantees (e.g. recall over a corpus with a planted near-duplicate, as line 508 does), and drop the 0.1 floor on random data.

Until then the ratchet will keep failing unrelated PRs at random; I re-ran the failed job on #3184 rather than touching the baseline, since adding it to scripts/ci-test-baseline.txt would hide a real regression in that file.

🤖 Generated with RuFlo

https://claude.ai/code/session_019xHM4rAH4aaShb4DTr1n6s