Tests were failed due to (now yanked) setuptools-scm v8.0.0, restarting tests

Thanks! I ran into the same issue managing to fetch 8.0.0 in the single hour it was online at condaforge... 😬

@dhomeier this looks almost good to go, are you planning on coming back to this?

This currently has an issue with the y limits of the selector 'snapping' to the bottom/x axis - to adapt an example @dhomeier shared with me:

import numpy as np
import matplotlib.pyplot as plt
from matplotlib.widgets import RectangleSelector

data = np.random.random((16, 16))
plt.ion()
fig, ax = plt.subplots()
ax.imshow(data)
selector = RectangleSelector(ax, lambda *args: None, interactive=True)
selector.extents = (5, 10, 6, 8)
plt.show(block=True)

The following happens:

For some reason the recording doesn't show the cursor but basically I first click on the lower right selection and move it slightly but it snaps to the bottom, then I touch the top right selector and the same thing happens.

@dhomeier - I've fixed the issue mentioned above in this PR to your branch: dhomeier#13

Another issue I ran into when testing this is that the region does not correctly follow the data when resizing the figure:

Thanks, seems I never properly set that example up with matplotlib on its own, but in astropy/regions#390 those operations are failing with a "negative height" set to the move or resized rectangle, where height was set from selector.extents[3] -selector.extents[2] (which have snapped to something like (5.2, 10.2, 14.4, 16.4) and then (5.2, 10.2, 15.4, 16.4) in the above example.

@dhomeier - I've fixed the issue mentioned above in this PR to your branch: dhomeier#13

Thanks, that does fix the original regions problem as well – no idea how I could not have looked at that!
Also still seeing some issues with the outline being offset.

Ok so the issue in #26833 (comment) is a little tricker as far as I can see - as the selector is defined in display coordinates, it does not follow the data when resizing the figure. We have to somehow cache the position of the selector in data coordinates so that we can change it in _on_resize. I'm not sure what the best way is to do this though, I've tried a couple of simple approaches which didn't work, I think the tricky bit will be dealing with the case with non-zero rotation.

Actually the issue I mentioned above to do with resizing is exactly what @jklymak mentioned in #21945 (comment)

Fundamentally, the issue is that we can't follow a resize since that might change the pixel aspect ratio and so a rotated rectangle would no longer be a rectangle. So maybe what we have here is fine, as it works provided that the user does not resize the window mid-selection.

@dhomeier - this needs to be rebased to remove the onselect=noop since that argument became optional.

This works nicely for me. I think there is no easy solution to the issue of the region changing in data coordinates when resizing the plot, there's fundamentally no way it will ever not change at all. Once this is rebased, I think it would make sense to proceed with reviewing and hopefully merging this!

That is remove them from all the test jobs, right?
Petty install failures on overwriting 2to3 on macOS_* are obviously unrelated.

Yes removing them from the tests. Just to check, did you lose the commit with my bugfix during the rebase?

Doh, thought I had rebased with that merged already!
Time to do some git-patch experiments...

During rebase two further issues showed up, addressed in the last two commits:

• Ellipse bounding box edges are defined relative to the centre, not the xy corner handle as in Rectangle (compare get_corners.
• For the selectors, @rotation.setter was using that corner handle as rotation point, while rotating interactively on the selector handles used the centre. For the downstream application in Adding support for rotating mpl selectors of Ellipse,RectanglePixelRegion astropy/regions#390 that turned out very inconvenient, as a combination of interactive and programmatic rotations would also introduce lateral movement of the widget, making it very difficult to reverse a rotation. Changing this behaviour might be more controversial, but I did not find any tests other than test_selector_rotate affected by this. The test updates hopefully make clear how I envision the functionality, but one could still consider adding a rotation_point property to RectangleSelector as well to make this configurable.

@story645, @ksunden we have completed our project in astropy/regions to a working state with this PR, so this should hopefully be ready for final review now.