The one issue seems to be that Numpy don't seem to have adopted the math.pxd file in the Numpy directory. I can find quite a few people using it, including Numpy themselves (https://github.com/numpy/numpy/blob/605e34577affa99fd49bbb6156843953b81d92b8/numpy/random/_common.pyx#L8) so unless I'm missing something, we may not be in a position to remove it

aargh. Sorry, we seem to have missed that: see numpy/numpy#25177. I guess Cython could drop the __init__.pxd file for now.

CI is passing other than the PyPy trashcan test, skipped in #5832.

About numpy.math.pxd:

@rgommers commented on elsewhere that perhaps the bettter solution is for cython to drop the file:

... [W]hat bothers me the most here is that the selection of functions included makes little sense. How would a user know what's in this namespace without trial and error?

It may be better for Cython to simply deprecate the numpy.math namespace and tell users to prefer libc.math and that if they have a specific issue to work around libc issues, simply do this in their own code:

cdef extern from "numpy/npy_math.h" nogil:
  # replace copysign with the function of interest here
  double copysign "npy_copysign"(double, double)

Is there any indication that numpy/math.pxd is in use?

Is there any indication that numpy/math.pxd is in use?

Looks like Scipy and Pandas both use it (just for NaN/INFINITY constants). scikit-learn looks to have used it until 2023 but it's probably been removed. A bunch of smaller projects also look to use it. I suspect not for anything irreplaceable, but I'm reluctant to cause that much disruption deliberately.

The SciPy usage was one very recently introduced line that I missed - I just opened a PR to remove it again. Pandas main branch looks clean too, no usages.

I think it would be good to at least approve the deprecation, conditional on code findable with GitHub code search being cleaned up first. Now that Cython 3 is kind of a "fresh" start with many much more disruptive changes than this one, it'd be nice to get rid of this.

I think this PR could be merged as-is. The use of math.pxd is a separate problem, and a little harder to solve, so separating the concerns and merging this would at least solve the immediate problem of duplicate files in two different repos.

Yeah I think I agree with this conclusion. I'll merge this for 3.1.

I mostly agree that we should drop numpy.math too. I'll make another PR for that but won't merge it unilaterally.