This looks close to right already, thanks @andfoy. Note that if you need to iterate on a CI job, it's usually best to do that on your own fork, to avoid running a ton of other jobs over and over. If you include a debug commit like rgommers/numpy@9e45f73, you'll trigger just this new job on your fork.

you'll trigger just this new job on your fork.

(on a new branch from which you didn't already make a PR)

(on a new branch from which you didn't already make a PR)

That makes sense, I'll do that then!

Regarding the test failure at scipy.misc.tests.test_doccer::test_decorator, multiline function docstrings seem to differ between Python 3.12 and 3.13:

def test_decorator():
    with suppress_warnings() as sup:
        ...
        @decorator
        def func():
            """ Docstring
            %(strtest3)s
            """

# Python 3.12
(Pdb) f.__doc__
' Docstring\n            %(strtest3)s\n            '

# Python 3.13
(Pdb) f.__doc__
'Docstring\n%(strtest3)s\n'

I don't know if this is an expected change or a regression in CPython, I took a look at the Changelog, but there is no mention to multiline strings besides python/cpython#112943. However, that issue doesn't seems to be the culprit here, since those changes were also backported into 3.12.2 (the output trace comes from a 3.12.3 instance)

From https://docs.python.org/3.13/whatsnew/3.13.html#other-language-changes: "Compiler now strip indents from docstrings." I bet that that's the culprit

Please feel to skip that test here and open a new issue about it, marked for 0.14.0

#19572 x-ref

Great to see the progress on this! Does that mean that there will be a development free-threaded wheel in https://anaconda.org/scientific-python-nightly-wheels/scipy/files at one point?

Asking mostly and out of curiosity and also with the goal one day of setting up a similar build for scikit-learn.

Yes, we should start building nightly wheels, at least for Linux x86-64.

@andfoy could you have a go at doing that as the next step here? I think it should look very similar to numpy/numpy#26512; if you start from copying SciPy's wheels.yml to free-threaded-wheels.yml and adapt from there, hopefully it'll work without new surprises.