FIX: Lock axes bounds in plot_joint to prevent shrinking on cursor motion (closes #14317)

[AkhilG-exe]

Why

evoked.plot_joint() figures shrink to zero once the cursor moves (or hovers) over them (closes #14317).

The joint plot is a layout="constrained" figure. The connection lines between the time series and the topomaps are drawn as ConnectionPatch(..., clip_on=False) artists added to the butterfly (ts_ax) axes. Matplotlib's constrained layout computes each axes' margins from Axes.get_tightbbox(for_layout_only=True), which includes every artist not fully clipped to the axes. Because these lines poke outside ts_ax and never move with it, every redraw (which cursor motion triggers through the interactive hover/_BlitManager path) measures a larger margin than the last, and the solver turns that growth back into a smaller axes — a feedback loop collapsing all axes to zero.

Reproduces with a plain sequence of fig.canvas.draw() calls:

redraws	1	2	3	10
butterfly axes height (before)	0.481	0.452	0.423	0.003
butterfly axes height (after)	0.507	0.507	0.507	0.507

What

• mne/viz/evoked.py: call con.set_in_layout(False) on the connection lines so constrained layout ignores them. This is the matplotlib-idiomatic mechanism for artists that intentionally leave their axes; rendering is unchanged and the initial layout is essentially the same.

• mne/viz/tests/test_topo.py: regression assertions added to the existing test_plot_joint — the largest axes' position must be unchanged after repeated redraws.

• doc/changes/dev/14319.bugfix.rst: changelog fragment.

• doc/changes/names.inc: new contributor entry.

Test plan

• pytest mne/viz/tests/test_topo.py mne/viz/tests/test_evoked.py → 25 passed

• test_plot_joint fails without the fix (regression) and passes with it

• TFR joint-plot tests (test_tfr_plot_joint, test_tfr_plot_joint_doesnt_modify) → 16 passed (TFR path was already safe: its ConnectionPatch is attached to the figure, not an axes)

• ruff check / ruff format --check clean on modified files

• Before/after visual verification: pre-fix renders drift by 122,985 pixels over 9 redraws; post-fix renders are bit-identical (0 pixels).

Disclosure per the AI-assistance policy: this change was developed with assistance from an AI coding agent (opencode), which diagnosed the constrained-layout feedback loop, while I implemented the one-line fix, and drafted the regression test. I reviewed and verified all changes. as well.

[CarinaFo]

Thanks for the quick bugfix, looks good. I verified the bug and the fix on Windows 11.

[drammock]

Thanks for submitting a fix. change looks good, test could be cleaner.

[drammock]

this line should be unnecessary; in a test like this the input data is known so the output of plot_joint (figure or list) is also known. Moreover: tests should be fast, so if the input data is generating multiple figures you should change the test so it only generates one (there's nothing about mag vs grad vs eeg that is expected to behave differently here, so testing with just one sensor type should be adequate).

[AkhilG-exe]

Thanks for the feedback, @drammock! I've updated the test to pass picks='grad' directly into plot_joint to isolate a single sensor type. This makes the test run significantly faster and allowed me to drop the redundant list-checking normalization entirely. Ready for another look!