I think we are hitting this in https://github.com/mne-tools/mne-python/actions/runs/21374406723/job/61528497819?pr=13613 / mne-tools/mne-python#13613:

_____________________ test_plot_raw_selection[matplotlib] ______________________
mne/viz/tests/test_raw.py:365: in test_plot_raw_selection
    fig._fake_click(xy, fig=sel_fig, ax=sel_fig.mne.radio_ax, xform="data")
mne/viz/_mpl_figure.py:2291: in _fake_click
    _fake_click(
mne/viz/utils.py:804: in _fake_click
    fig.canvas.callbacks.process(
/opt/hostedtoolcache/Python/3.14.2/x64/lib/python3.14/site-packages/matplotlib/cbook.py:365: in process
    func(*args, **kwargs)
/opt/hostedtoolcache/Python/3.14.2/x64/lib/python3.14/site-packages/matplotlib/widgets.py:193: in wrapper
    return func(self, event)
           ^^^^^^^^^^^^^^^^^
/opt/hostedtoolcache/Python/3.14.2/x64/lib/python3.14/site-packages/matplotlib/widgets.py:1113: in _clicked
    *self._frames.contains(event)[1]["ind"],
     ^^^^^^^^^^^^
E   AttributeError: 'RadioButtons' object has no attribute '_frames'

Where our code does

    fig.canvas.callbacks.process(
        kind,
        backend_bases.MouseEvent(
            name=kind, canvas=fig.canvas, x=x, y=y, button=button, key=key
        ),
    )

I notice there is no test changed so I wonder if it's worth adding some minimal test that would have caught this sort of error to prevent future regressions? This for example fails for me with that same traceback on 4ca5e0c (but I'm sure it could be improved!):

from matplotlib import widgets, backend_bases
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
fig, ax = plt.subplots()
button = widgets.RadioButtons(ax, ["Test Button"])
fig.canvas.callbacks.process(
    "button_press_event",
    backend_bases.MouseEvent(
        name="test", canvas=fig.canvas, x=200, y=200, button=1, key=None
    ),
)

Adding a test is indeed a good idea. I added it before the fix, so it'd be easier to verify it works well.

Thanks for the comments :).

Did you recheck the test before the change? It seems to pass if I run all of test_widgets.py, but fails if running only the one test. I'm not sure what's going on there.

PS, there's a typo in the first commit message.

Did you recheck the test before the change?

Yes, I checked before the last commit of the PR and it fails there, and after the last commit of the PR it succeeds.

It seems to pass if I run all of test_widgets.py, but fails if running only the one test. I'm not sure what's going on there.

I have no idea why is that. What is the error you get? I tested it myself now and it works as explained above.

PS, there's a typo in the first commit message.

Do you mean it should be: widets: test _Buttons' callbacks and not widets: test _Buttons callbacks?

It seems to pass if I run all of test_widgets.py, but fails if running only the one test. I'm not sure what's going on there.

I have no idea why is that. What is the error you get? I tested it myself now and it works as explained above.

No error, I mean it incorrectly passes before the fix if run together.

PS, there's a typo in the first commit message.

Do you mean it should be: widets: test _Buttons' callbacks and not widets: test _Buttons callbacks?

It should be widgets.

I've narrowed it down to if you run test_span_selector_animated_artists_callback first then the test erroneously passes before the fix is made:

$ pytest lib/matplotlib/tests/test_widgets.py -k 'test_span_selector_animated_artists_callback or test__buttons' -v
========== test session starts ==========
platform linux -- Python 3.11.4, pytest-7.4.0, pluggy-1.2.0 -- python3.11
cachedir: .pytest_cache
rootdir: .../matplotlib
configfile: pyproject.toml
plugins: cov-4.1.0, rerunfailures-12.0, timeout-2.1.0, xdist-3.3.1
collected 126 items / 123 deselected / 3 selected                                                                                                                                                                  
lib/matplotlib/tests/test_widgets.py::test_span_selector_animated_artists_callback PASSED [ 33%]
lib/matplotlib/tests/test_widgets.py::test__buttons_callbacks[RadioButtons] PASSED        [ 66%]
lib/matplotlib/tests/test_widgets.py::test__buttons_callbacks[CheckButtons] PASSED        [100%]
========== 3 passed, 123 deselected in 3.25s ==========

I'm going to open a separate issue for that as it's a deeper, but unrelated, testing problem.

Edit: can you also modify the test to call on_clicked with a Mock and assert that it was called after the mouse event? You may have to tweak the coordinates a bit.

PS, there's a typo in the first commit message.

Do you mean it should be: widets: test _Buttons' callbacks and not widets: test _Buttons callbacks?

It should be widgets.

OK 😄 done.

I'm going to open a separate issue for that as it's a deeper, but unrelated, testing problem.

OK 🙏 . For the record, I also tried to debug it a little bit, and this seems to unrelated to the get_ax pytest fixup, meaning the following diff:

diff --git i/lib/matplotlib/tests/test_widgets.py w/lib/matplotlib/tests/test_widgets.py
index 8b421aab33..02e653e1e8 100644
--- i/lib/matplotlib/tests/test_widgets.py
+++ w/lib/matplotlib/tests/test_widgets.py
@@ -1220,10 +1220,16 @@ def test_check_button_props(fig_test, fig_ref):
 
 
 @pytest.mark.parametrize("widget", [widgets.RadioButtons, widgets.CheckButtons])
-def test__buttons_callbacks(ax, widget):
+def test__buttons_callbacks(widget):
+    fig, ax = plt.subplots()
     """Tests what https://github.com/matplotlib/matplotlib/pull/31031 fixed"""
     button = widget(ax, ["Test Button"])
-    MouseEvent._from_ax_coords("button_press_event", ax, (200, 200), 1)._process()
+    fig.canvas.callbacks.process(
+        "button_press_event",
+        MouseEvent(
+            name="test", canvas=ax.figure.canvas, x=200, y=200, button=1, key=None
+        ),
+    )
 
 
 def test_slider_slidermin_slidermax_invalid():

Doesn't fix the issue.

Edit: can you also modify the test to call on_clicked with a Mock and assert that it was called after the mouse event? You may have to tweak the coordinates a bit.

That wasn't trivial, but indeed it was a good idea.

I'm going to open a separate issue for that as it's a deeper, but unrelated, testing problem.

I feel like the current test sort of hides this problem, so I wasn't sure whether a comment should be written to explain that.

I'm going to open a separate issue for that as it's a deeper, but unrelated, testing problem.

I feel like the current test sort of hides this problem, so I wasn't sure whether a comment should be written to explain that.

Well, that was sort of intentional. If the exception occurs, then the callback isn't called and the test fails. This makes the test resilient to whether it is run before or after the GUI test, since it doesn't matter if the exception is actually fully raised.