The 6 failures here are similar to this one (See the S-E bar for on both and one is filled all the way to the circle, the other one is not):

I'll increase the tolerance on a case by case basis to avoid to ensure we can get these to pass but without losing meaningful difss.

This one is ready for review. When merged I'll unpin rebase #191 and unpin matplotlib to ensure we catch all the differences in the new version.

I just hope this kind of tests won't break too often.
Let's give it a try.

Lgtm

I just hope this kind of tests won't break too often.

I have the same concern but cartopy has been using pytest-mpl for a while and it seems quite robust. My local tests showed the "pacman" problem reported in #190 nicely. However, if things are too brittle we can dial back and use just a bare axes for the test with a allow_failure flag.

Let's give it a try.

👍