Fix zppy testing

[xylar]

Fix panels dropped from projection comparison plots under matplotlib 3.11

Summary

Under matplotlib 3.11, roughly half the plots produced by an MPAS-Analysis run come out visibly broken: an entire map panel is cropped out of the figure and the subplot titles are missing. This is not a plotting-style regression that can be waved away as cosmetic, and it is not a change in the underlying data — the affected figures are genuinely unusable. This PR fixes it with a one-line change, and adds some unrelated test-hygiene cleanups that surfaced along the way.

The problem was originally reported against zppy's weekly integration testing, where 852 of 856 MPAS-Analysis images for comprehensive_v2 failed the image comparison: E3SM-Project/zppy#845

Root cause

plot_projection_comparison() in mpas_analysis/shared/plot/climatology_map.py was the only one of the three comparison plotting functions that called ax.set_title() without an explicit y. That matters because matplotlib only auto-positions a title when y is None, and that auto-positioning inspects the surrounding artists in order to place the title clear of them — including the cartopy gridline labels, whose bounding boxes are non-finite while the figure is being drawn.

The title therefore ended up at y = inf, which left the entire map axes with a non-finite tight bounding box. The subsequent savefig(bbox_inches='tight') in mpas_analysis/shared/plot/save.py then unions only the axes whose bounding boxes are finite — which is just the colorbars — so the saved figure began at the first colorbar and the leading map panel fell outside the image entirely.

Instrumenting the three-panel sea ice concentration plot on matplotlib 3.11 shows this directly, where ax0/ax2/ax4 are the map panels and ax1/ax3/ax5 are their colorbars:

ax0 title_pos=(0.5, inf)  tightbbox_finite=False  bbox=Bbox(x0=nan, y0=222.4, x1=nan, y1=inf)
ax1 title_pos=(0.5, 1.0)  tightbbox_finite=True   bbox=Bbox(x0=1505.7, y0=269.7, x1=1674.3, y1=1215.3)
ax2 title_pos=(0.5, inf)  tightbbox_finite=False  bbox=Bbox(x0=nan, y0=222.4, x1=nan, y1=inf)
...
FIG tightbbox: Bbox(x0=1505.7, ...)

The figure's bounding box starts at x0=1505.7, which is exactly the left edge of the first colorbar rather than the left edge of the first map. After the fix every axes is finite and the figure bounding box starts at x0=550.0, the left edge of the first map panel.

plot_polar_comparison() (y=1.06) and plot_global_comparison() (y=1.02) already passed an explicit y and were never affected, which is why the bug is confined to the projection maps.

The fix

Pass an explicit y in plot_projection_comparison() so that matplotlib does not auto-position the title, matching what the other two functions already did. This is seven lines including the explanatory comment.

This is the same class of problem addressed in E3SM-Project/polaris#635 and MPAS-Dev/compass#972, though the remedy differs: Polaris dropped bbox_inches='tight' in favor of constrained layout, and Compass switched to requesting only the gridline labels it wanted. Requesting only the wanted labels was tried here first and did not fix it — the title still landed at y = inf — so the minimal change that actually works is the explicit y. That also avoids raising the cartopy requirement to >=0.22, which list-valued draw_labels would have needed.

Verification

Two pixi environments were built that differ only in the matplotlib version — python 3.14.7, numpy 2.5.2, cartopy 0.25.0, pillow 12.3.0 and freetype 2.14.3 are identical in both — and a full zppy comprehensive_v2 MPAS-Analysis run was done in each against v2.LR.historical_0201 on Chrysalis.

Standalone reproducer, the antarctic_extended sea ice concentration comparison:

	matplotlib 3.11.1	matplotlib 3.10.9
before the fix	2620x1195 (broken)	3575x1242
after the fix	3576x1242	3575x1242

Across a full 856-image run, comparing matplotlib 3.11.1 against 3.10.9 before the fix, 462 images (54%) differed by more than 10 pixels in at least one dimension — that is, broken layout rather than a text-metric shift — split 284 sea ice and 178 ocean. The remaining 372 differ by 10 pixels or fewer, which is the ordinary matplotlib 3.11 text-metric change and affects essentially every figure any version of this code produces.

After the fix, all 462 previously-broken images render correctly. Every one of the 178 ocean failures is resolved, and sea ice drops from 284 to 70. Those 70 remaining were inspected and are cosmetic: the height is identical in all 70 cases and only the width differs, by 11 to 16 pixels out of about 1620, and they are all seaice_{area,volume}tend{thermo,transp} plots, whose long titles let text-metric drift accumulate past the 10-pixel cutoff used here. Comparing the fixed run against the broken run, exactly 462 images differ — precisely the set that was broken — and the remaining 394 are unchanged in size, so the fix touches nothing else.

Also in this branch

These are independent cleanups that came out of getting the above tested, and can be split out if preferred.

• distutils removed from the test fixtures. distutils left the standard library in Python 3.12, so from distutils import dir_util only kept working through the shim setuptools installs. In an environment without setuptools, collecting anything under mpas_analysis/test failed outright. shutil.copytree(..., dirs_exist_ok=True) is the direct replacement.
• A test that the four dependency lists stay in sync. Dependencies are declared in dev-spec.txt, pixi.toml, ci/recipe/recipe.yaml and pyproject.toml, and nothing kept them consistent. The new tests check that the three conda lists agree exactly, that conda build pins match, that pyproject.toml agrees on every shared dependency with the set it legitimately omits pinned to a documented CONDA_ONLY list, that requires-python matches the python constraint everywhere, and that the dev and docs groups agree. Perturbing each of the four files was verified to be caught by the expected test.
• The test suite no longer writes into the repository. Running pytest mpas_analysis/test left inset_point.png, inset_region.png, inset_transect.png and PET0.RegridWeightGen.Log in the repository root. The PNGs came from bare relative filenames in test_inset.py; the ESMF log is written by the ESMF_RegridWeightGen subprocess that pyremap launches, which gets no cwd and so cannot be redirected from here. An autouse fixture now runs each test in its own temporary working directory, which covers the subprocess case too. test_inset.py additionally writes to tmp_path explicitly, closes its figures, and now actually asserts that the plot was written — previously those three tests could not fail.

Note for downstream image comparisons

Any project comparing MPAS-Analysis output against stored reference images will need to regenerate its baselines when matplotlib 3.11 arrives, independently of this fix. matplotlib 3.11 changed text metrics slightly, so every figure saved with bbox_inches='tight' comes out one or two pixels different in size and fails an exact-pixel comparison. That accounts for the 43% of images that were shifted but not broken.

Checklist

• Testing comment in the PR documents testing used to verify the changes

[xylar]

Testing

The problem with matplotlib 3.11.1 shows up in:
https://web.lcrc.anl.gov/public/e3sm/diagnostic_output/xasaydavis/zppy_weekly_comprehensive_v2_www/mpl_repro_20260823_dev/v2.LR.historical_0201/mpas_analysis/mvo/ts_1980-1984_climo_1980-1984/sea_ice/index.html

It looks right with the fix from this branch in:
https://web.lcrc.anl.gov/public/e3sm/diagnostic_output/xasaydavis/zppy_weekly_comprehensive_v2_www/mpl_repro_20260823_fix/v2.LR.historical_0201/mpas_analysis/mvo/ts_1980-1984_climo_1980-1984/sea_ice/index.html

[xylar]

Broken:

Fixed: