Some of this is #27624, but there are a number of image comparison failures (and one array comparison failure) that are caused by something else (that I haven't tracked down yet)... very low RMS, but something changed.

My entirely unproven (as yet) suspicion is that it is numpy 2.0 related changing some behavior.

Okay, I have bisected at least the majority of the errors to:

numpy/numpy@197e61c

(From numpy/numpy#23912)

Which are related to NEP50

I'm a little iffy on the timing, as that was merged in October, but we only started seeing the failure within the last 2-3 weeks.

That said, the specific commit is changing the default value which can be overridden by env var, and changing back to "legacy" mode (i.e. disabling NEP 50 behavior) does in fact pass the tests.

However, I think it is clear that at least as far as the pie changes go, it is a negligible result of different dtype rules in intermediate calculations and we can just add a tolerance to the tests.

There are two additional tests that I've not yet fully tracked down:

• FAILED lib/matplotlib/tests/test_basic.py::test_override_builtins
  • Likely related to numpy namespace changes, probably needs a slight adjustment to account for that change
• FAILED lib/matplotlib/tests/test_colors.py::test_scalarmappable_to_rgba[False] - AssertionError:
  • Appears to be caused by the same numpy change, may just need to be np.assert_almost_equal instead of assert_equal, as there is a slight amount of floating point rounding going on there.

Indeed the test_override_builtins is related to numpy/numpy#25086

The extra things that are affected are pow and bool, I think I will add those to the list of things we set back as the builtins version at the bottom of pylab.py, and maintain current behavior there.