The only CI failures are unrelated (due to gh-27548). The BLAS detection in those two failing jobs worked fine, so this looks about ready.

Will add another fix and retest.

Tagged this for backport, I will be making at least one more 2.1 release.

Is this based off current main? The macos failure might go away with a rebase.

Yeah it's up-to-date, only a few commits behind. The macOS conda-forge failure is coming in with the AIX linker change in the vendored Meson it looked like, the job is passing on the first commit in this PR EDIT: looks like it was flaky - it's failing again. Still puzzled by it though, not sure what's going on there.

I'll drop the change for now - the primary goal here is to update to Meson 1.5.x and get free-threading on Windows going.

Ugh, the macOS failure is a flake that just happened to also touch on icc. It shows up on gh-27544 from 3 days ago as well, and also on several other issues from years ago (e.g., gh-17635). None of the other issues have an explanation.

I see the test run is using xdist with two workers. Is there something in the test that is not thread safe? Maybe some other test is importing and changing sys.modules in parallel with this one?

The failures are in numpy.distutils which is deprecated. Can we avoid testing whether any of its API has changed?

macOS failure is a flake

The reason I ask is that it was common until a cache update. See #27532.

The reason I ask is that it was common until a cache update.

Hmm, not impossible, although I think my initial PR included that fix already (it was anchored on a commit only 2 days old). That said, your "wrong setuptools version" is quite plausible.

The failures are in numpy.distutils which is deprecated. Can we avoid testing whether any of its API has changed?

An easy way to do that would be to bump this CI job to use Python 3.12 (assuming nothing else is broken there). I'll check that.

Did you mean to change this?

Note that the problems seemed to start with a new release of 3.11.

Did you mean to change this?

Yes indeed. Done in gh-27555.

Okay, this is finally ready now. I updated the PR description with what it adds/fixes. Should be back-portable for 2.1.3.

Thanks Ralf.