BLAS thread count alone changes scanpy Leiden clusters on one machine

Same env (scanpy 1.12.5, OpenBLAS 0.3.34), same seed, same machine: changing only OPENBLAS_NUM_THREADS/OMP_NUM_THREADS from 8 to 1 changed the PBMC3k Leiden partition in 10/10 seeds (ARI 0.63-0.996, median 0.98; one seed moved 30% of cells with the same cluster count). Each thread setting was bit-repeatable, and NUMBA_NUM_THREADS had no effect. The path: the BLAS reduction order changes the last digits of the PCA, 0-2 kNN ties flip or weights move about 1e-3, and Leiden amplifies that. For reproducible single-cell clustering, record and pin the BLAS thread count along with the package versions. Otherwise the result depends on how many cores the machine or scheduler gives you.