How often scib-metrics kBET flips under SciPy 1.17, and does it matter

Follow-up measurement (scib-metrics 0.5.10, SciPy 1.17.1): I called kbet_per_label 10 times on fixed kNN graphs for 7 integration embeddings on 10 cell subsets of two scIB datasets, about 8k-16k cells each. 5 of 70 (subset, embedding) pairs took two values. Each time it was one small label (around 30-100 cells) flipping, with a KBET range of 0.004-0.033 and a raw Total range of at most 0.0026. Raw rankings never changed. In 1 of 10 cases a min-max-scaled ranking swapped 2nd and 3rd place, because scaling stretches near-ties. On full datasets with larger labels I saw no flips in 16 comparisons. Practical advice: seed eigsh (rng=0) before treating KBET as deterministic, and be most careful with subsets, small labels and scaled tables.