Empirically, I think I agree that the changed behavior is better (at least with the main example provided here, and don't have any reason to suspect it is special).

However, the main hangup with the previous PR appears to be a request for an explanation of why the offset math works out this way in the comments (phrased as a request for a "brief proof"), which I do not think this PR provides at this time.

If I read the code correctly the primary idea is that you are adjusting the anchor point by the applied rotation. The comment that exists "Standard 2D rotation transformation" does not in my opinion adequately explain it's connection to the anchor point. The change in sign in the rounding line is also mildly confusing to me (not saying it is wrong, just not sure why it changed)

Finally, yes, before this can be merged, the tests will have to be updated so they are passing. I don't love the quantity (or magnitude) of changes, as we do try to have stable outputs. However, correctness is certainly a good reason to change them.

I'm not sure why the one test changes the line at all... if it was positioning/spacing of the axes I'd have expected the axis frame to have differences in the diff output. Further confused why it has a gradient of magnitude of the change.

Thank you for the comment.

I've updated the code to hopefully adequately explain the logic behind the changes.
Regarding the failing tests, should I include the updated images in this pull request or is there some other procedure for how to update them (assuming that this is the correct solution and that they should be changed)?

What's the best thing to do about the tight_layout test? I don't see why the image is changed but at the same time I see no issues with other plots that I think would be present if my changes did something weird with the plotting logic.

I'm not sure why the one test changes the line at all... if it was positioning/spacing of the axes I'd have expected the axis frame to have differences in the diff output. Further confused why it has a gradient of magnitude of the change.

That's a tight layout test; when the text moves a little bit, so does everything else. It's fine so long as the text movement is expected.

Regarding the failing tests, should I include the updated images in this pull request or is there some other procedure for how to update them (assuming that this is the correct solution and that they should be changed)?

Yes, you should update the test images in the same PR. You can usetools/triage_tests.py to check them.

Yes, you should update the test images in the same PR. You can usetools/triage_tests.py to check them.

I've now updated the images to the new ones. Thank you for the information.

Can you add a new image test which is the reproducer from this issue? Although any future changes to this will clearly cause a bunch of failures, having one very clear test case is still very valuable.