Minimum Python available there is 3.10 now, so bumped that up. It looks like pycairo and PyGObject don't have wheels. PyQt5 and wxPython4 have wheels, but I haven't tried allowing the others that were blocked on macOS yet.

We do seem to be getting a lot more failures here than we did when trying Cirrus in #24597.

I've de-conditionalized all optional GUI dependencies so I can verify whether they should be skipped or not.

The results are the following:

Most are available everywhere.

The PyGObject 3/4 split doesn't have any effect, as the dependencies are the same either way (the only difference is the native dependency.) Only PyGObject 4 on macOS 14 arm64 is missing wheels, but appears to be available regardless after building them.

PySide2 on Python 3.12 installs a wheel, but crashes. I believe PySide2 is no longer supported so they didn't bother fixing it for Python 3.12, and because it's some kind of universal thing, it still installed. On the other hand the M1 was new enough that there were no wheels for it.

I rebased #24597 for comparison, and it doesn't seem to fail nearly as many tests. I can't see any tolerance issues at all. For a reason I don't understand, xfail appears to be broken on both.

Ah, looking at the pytest 8 changelog, xfails are now fully reported with backtraces. This seems far too noisy, and we may have to disable that.

(mpl-dev) tcaswell@seven matplotlib %  python -c 'from matplotlib.backends._macosx import print_screen_info; print_screen_info()'                                                                                                                                                                                      macos-m1
mainScreen: Acer B286HK (3840.000000x2160.000000)+0.000000+0.000000 @8
deepestScreen: Acer B286HK (3840.000000x2160.000000)+0.000000+0.000000 @8
Available screens: 1
    0: Acer B286HK (3840.000000x2160.000000)+0.000000+0.000000 @8
(mpl-dev) tcaswell@seven matplotlib %  python -c 'from matplotlib.backends._macosx import print_screen_info; print_screen_info()'                                                                                                                                                                                      macos-m1
mainScreen: (unknown)
deepestScreen: (unknown)
Available screens: 0
(mpl-dev) tcaswell@seven matplotlib %  python -c 'from matplotlib.backends._macosx import print_screen_info; print_screen_info()'                                                                                                                                                                                      macos-m1
mainScreen: (unknown)
deepestScreen: (unknown)
Available screens: 0

That is the result for me running over ssh with:

• account logged in and a screen connected (but locked)
• account logged out and a screen connected
• account logged out and no screen connected

The account I was logged in via and the account I was ssh'd to were the same. My guess is that being logged in as a different user on the screen would still result in no screens being reported via ssh (but I have not tested that if we care I can).

We should definitely adapt this and use it the same way we use the "do we have a screen" check on linux to auto fallback to Agg.

OK, I removed the debugging, and pushed the screen check to #27761.

We decided to xfail the test_other_signal_before_sigint test, because sometimes seems fixed, and sometimes doesn't, regardless of the commit above, and @ksunden reproduced once, but not much luck after that.

Also, contrary to our call @ksunden / @tacaswell, I decided to keep the native arm64 build here. In fact, the cross-compiled one is brokenly-named as well, so we need to fix it one way or the other, and might as well do it now.

Because this fixes the wheel names, this is important to get in to 3.9 before the rc.

I only half understand the comment in the svg backend.

I made some edits to those comments.