umap-learn <=0.5.11 with numba >=0.62 gives sigma=inf, corrupting scanpy kNN graphs

umap's smooth_knn_dist is jitted with fastmath=True and tests hi == NPY_INFINITY. With numba 0.62.0 or later (checked up to 0.68.1) that branch misbehaves, the binary search returns sigma=inf, and every kNN membership for those cells becomes 1.0. On the scanpy PBMC3k tutorial PCA this hit 15% of cells (k=10), 29% on scanpy's bundled pbmc68k_reduced, and 13-16% on paul15. Total connectivity weight rose 24%, and there is no error or warning. A fresh pip install scanpy resolved between 2025-09-18 and 2026-04-08 got this combination. umap-learn 0.5.12 fixes it by using NPY_FLOATMAX as the sentinel; numba 0.61.2 also avoids it. Quick check: np.isinf(umap.UMAP(transform_mode='graph').fit(X)._sigmas).mean() should be 0. On PBMC3k, Leiden clusters moved only slightly beyond seed noise (ARI 0.93 vs 0.95), so downstream results can look fine while the graph is wrong.