numba >=0.62 fastmath=True folds x == inf and np.isinf(x) to False
Minimal test on CPython 3.12. With @njit(fastmath=True), f(np.inf) returns True for x == np.inf and for np.isinf(x) on numba 0.61.2 (llvmlite 0.44, LLVM 15), and False for both on 0.62.0 (llvmlite 0.45.1, LLVM 20) and 0.68.1. A loop that uses hi = np.inf as an 'unset' sentinel never sees it as unset. fastmath=True includes the 'ninf' flag (assume no infinities), and the newer LLVM uses it to fold these checks to constants. Passing the flag set without ninf, fastmath={'nnan','nsz','arcp','contract','afn','reassoc'}, made x == np.inf correct again on 0.62 and 0.68. Real casualty: umap-learn <=0.5.11 smooth_knn_dist, where 95% of sigmas on sklearn digits became inf. If jitted code uses inf (or NaN, which 'nnan' covers) as a sentinel, drop those flags or use a finite sentinel such as np.finfo(dtype).max.
Self-check of the NaN remark: with @njit(fastmath=True), np.isnan(np.nan) and x != x on a NaN both return False on numba 0.61.2 and on 0.62.0. So NaN checks were already folded away before 0.62 (via nnan); only the inf case is new in 0.62. Same fix: leave 'nnan' out of the fastmath flag set when the code checks for NaN.