SciPy 1.17/1.18 removals; sph_harm -> sph_harm_y swaps arguments silently
Signature diff plus calls, SciPy 1.16.3 -> 1.17.0 -> 1.18.1. Everything below gave a DeprecationWarning on 1.16.3. Gone in 1.17.0 (AttributeError): linalg.kron (use np.kron), special.sph_harm, lpn, lpmn, clpmn, stats.find_repeats, distance.kulczynski1 and sokalmichener; ttest_ind(permutations=, random_state=) raise TypeError (use method=stats.PermutationMethod(...)). Gone in 1.18.0 (TypeError): disp= in linalg.logm, signm and sqrtm, sqrtm(blocksize=), fmin_l_bfgs_b(iprint=, disp=), nnls(atol=). The trap is the sph_harm replacement: sph_harm(m, n, azimuth, polar) equals sph_harm_y(n, m, polar, azimuth), so both the order and the angle convention swap. On 1.16.3, renaming only the function, sph_harm_y(1, 2, 0.7, 0.4), returned 0j without a warning (it reads n=1, m=2, |m| > n), while the old call gave -0.212-0.179j. legendre_p_all(2, 0.5)[0] matches lpn(2, 0.5)[0].
Confirmed on SciPy 1.16.3, 1.17.0, 1.18.0 and 1.18.1 (CPython 3.12.16, uv wheels). With DeprecationWarning turned into an error, every call listed raises it on 1.16.3: kron, find_repeats, kulczynski1, sokalmichener, ttest_ind(permutations=/random_state=), disp= on logm, signm and sqrtm, sqrtm(blocksize=), fmin_l_bfgs_b(iprint=/disp=), nnls(atol=). On 1.17.0 the attributes give AttributeError and the ttest_ind keywords TypeError, while the 1.18 group still only warns. On 1.18.0 and 1.18.1 that group gives TypeError. sph_harm(1, 2, 0.7, 0.4) = sph_harm_y(2, 1, 0.4, 0.7) = -0.2119-0.1785j, and the rename-only call returns 0j with no warning on every version tested. legendre_p_all(2, 0.5)[0] = lpn(2, 0.5)[0] = [1, 0.5, -0.125]. One limit: for lpmn and clpmn I checked only that they are gone in 1.17.0, not their 1.16.3 warnings or replacements.