FIX: Handle hidden annotations during deletion in mpl plot

[DerAndereJohannes]

Reference issue

Fixes #13511.

What does this implement/fix?

This pull request fixes an issue in which removing annotations from the raw matplotlib figure would remove additional unintended annotations.

The issue with the current live code is that it assumes that the visible annotations are all grouped together in a continuous block starting from offset. However, this logic does not hold up.

For example, if is_onscreen is [True, False, True], then the offset is 0. Therefore zorders[0:2] get the visible_zorders instead of index 0 and 2.

The fix simply masks the zorders using the is_onscreen to assign visible_zorders to the correct indices.

Additional information

I have just implemented the fix here without a new regression pytest. Let me know how I should implement one if it is needed. Thank you!

[larsoner]

Code looks good to me. Any way you could update a test that would fail on main but pass on this PR? Maybe somewhere in

mne-python/mne/viz/tests/test_raw.py

Line 768 in 8dd298f

def test_plot_annotations(raw, browser_backend):

And you can look at _annotation_helper to see some logic/magic to interact with the window using code.

It'll be a pain (harder than implementing the actual fix!) but I fear without this we can too easily regress in the future

[DerAndereJohannes]

No worries, I will give it a shot! Thank you for the hints.