MAINT: version bounds for 1.11.0rc1

[tylerjereddy]

• the version bounds are easy to mess up--the reference conversation from the last release is linked below:
  MAINT: version bounds for 1.10.0rc1/relnotes fixes #17569

• the adjustment of version strings to include i.e., rc1 is normally left to the REL commit proper during the release process

Other vague things on my mind include:

• doc version switcher adjustments (I think I can get away with delaying those until later in the process)
• NumPy 2.0 coming sooner than later--I'm not sure this really requires any special action at the time of writing--I suspect producing a helpful error message with a major version change is the behavior we want for now
• this branch may eventually need to support Python 3.12

[tylerjereddy]

There's a segfault on alpine in CI: scipy/linalg/tests/test_decomp.py", line 996 in test_random_complex didn't check if that has happened on main recently yet.

The prerelease failure is working as I currently intend it: meson.build:57:2: ERROR: Problem encountered: SciPy requires Cython < 3.0. If we plan to backport support for 3.x (or are confident enough as-is), I can remove that I suppose...

[ilayn]

you can ignore that segfault. I couldn't pinpoint the exact time but when multiple workers are fired, it runs out of memory somehow in the linalg tests. I already reduced the numbers on certain tests but I think it is dying because there are too many tests in that module ~3k LOC file.

[h-vetinari]

I think Cython 3 compatibility would be good to have also on the 1.11 branch. We should be pretty close already, hopefully any relevant patches would be small and easy to backport.

[rgommers]

The Cython <3.0 looks correct to me, we are likely to see more breakage there so should exclude it. The Python 3.12 support could go either way; I suggest allowing it already because it likely works and early adopters are starting to build for it from source.

[tylerjereddy]

Revised to allow Python 3.12 support.

[rgommers]

I think Cython 3 compatibility would be good to have also on the 1.11 branch. We should be pretty close already, hopefully any relevant patches would be small and easy to backport.

To expand on my answer: we add a <3.0 upper bound because we want pip install scipy with scipy being built from source to work. So we add upper bounds to all dependencies to achieve that. It does not mean that we won't be able to backport fixes for 3.0 compat from main to 1.11.x, and if Cython 3.0 gets released and we do a bugfix release from the 1.11.x branch after that, we can enable it just fine.

[h-vetinari]

To expand on my answer [...]

This is what I meant as well. 👍

[rgommers]

This doesn't actually work exactly as intended unfortunately, because .version_compare does not deal with pre-releases. So a 3.0.0b3 version will choke on this check. Now, that may be intended since those pre-releases aren't really supported. However, the 3.0 release is expected soon and it actually does work. So I think we should remove this check again. We'll keep it in pyproject.toml, so by default users who pip install scipy and get a from-source build will get the latest 0.29.x version. But if they are explicitly using 3.0, let's let them do that.