The debug crash I'm not sure about, seems unlikely to be related but it hasn't happened on other PRs recently. The Cygwin failure seems relevant, lots of BlockingIOError: [Errno 11] Resource temporarily unavailable on f2py tests. Unclear why.

The debug job has now crashed twice:

./tools/travis-test.sh: line 82:  4665 Aborted                 (core dumped) 

and the Cygwin job also doesn't pass. So there is more wrong than just missing variables probably. Let's see if a revert is happier (see gh-23067).

Cygwin just had a release, on Jan 17, and then another bugfix one last night (Jan 22): http://mirrors.kernel.org/sourceware/cygwin/x86_64/. So that's probably the cause here. And the debug build is due to the recent change in gh-23026. So both should be unrelated. The exact same failures show up on a straight revert of the problematic PR (gh-23067), which confirms that. I've restarted the Cygwin job one more to verify whether it passes now (it did on another PR an hour ago).

Okay, CI jobs are consistently failing. I'm going to ignore that, merge this PR, and open follow-up issues for Cygwin and the debug job. Having Fortran support for numpy.distutils being completely broken is the worse problem here.

Follow up issues fixed: gh-23070, gh-23071

The Cygwin failure seems relevant, lots of BlockingIOError: [Errno 11] Resource temporarily unavailable on f2py tests. Unclear why.

That's usually fork() failures, which are more likely to be due to CI setup than tests. I started rebasing numpy and test dependencies a few months ago, so my suspicion is some change in the CI runner. Unfortunately, nothing in the Jan 19 change list looks relevant (i.e. BLODA, usually antivirus or other software injecting itself into every process). More discussion on #23073.