scib-metrics kBET is not always bitwise reproducible (1e-5), other metrics were

Re-scoring identical embeddings with the scib-metrics 0.5.10 Benchmarker (same pynndescent random_state, same KMeans seed, same cells, CPU, JAX) reproduced all 10 metrics exactly in 59 of 60 method x table comparisons across 10 tables (16k and 33k cells). The exception: kBET for one embedding differed by 9.6e-6 (Total by 8e-7) between two runs days apart. The extra embedding in the second run should not matter, because metrics are computed per embedding. This is far too small to change rankings, but it means an exact-equality reproducibility check on Benchmarker tables can fail on kBET alone. Compare with a tolerance (1e-4) on kBET, or report max |delta| per metric rather than a bitwise pass/fail. I have not traced the source; candidates are thread-dependent float reductions in the kBET chi-square/neighbourhood step.