test(plotting): replace bilinear-resize fallback; fix dpi propagation (#1327)

[Marius1311]

Closes #1327.

Problem

tests/_helpers.py::resize_images_to_same_sizes ran before every compare_images, bilinearly resizing the rendered figure to the baseline's pixel size whenever they differed. A size mismatch is itself a signal that something changed (matplotlib/layout/DPI/backend), and resizing both:

• masked real regressions — a genuinely different layout was squashed to match, then passed under the RMS tolerance, and
• manufactured fake diffs — interpolation blurs the image before the RMS comparison.

What changed

1. Harness: resize fallback → size tolerance + crop

(tests/_helpers.py, tests/test_plotting.py)

• Remove resize_images_to_same_sizes.
• Add assert_image_sizes_close — a relative size tolerance (SIZE_RTOL = 0.1, the size-axis analogue of STRICT_TOL): fail loudly on gross size divergence, absorb sub-pixel bbox_inches="tight" rounding drift across platforms / matplotlib versions.
• Add crop_images_to_common_size — crop (no interpolation) so compare_images gets identical dimensions without blurring. The RMS comparison then runs on real, non-interpolated pixels.

This is consistent with #1328's philosophy (robustness via low DPI + tolerance, not pinning) rather than a strict size equality, which would flag ~17 sub-pixel-jitter tests — including baselines #1328 just regenerated.

2. dpi propagation fix

(src/cellrank/_utils/_utils.py, src/cellrank/estimators/terminal_states/_term_states_estimator.py)

Removing the resize exposed that plot_macrostates rendered ~2.6× too large: _plot_discrete and the _plot_continuous non-same-plot branch popped dpi and discarded it (the # handled at figure level comment was never implemented), so sc.pl.embedding rendered at scanpy's default ~100 dpi instead of the requested 40.

save_fig now accepts an optional dpi forwarded to fig.savefig, and both embedding paths pass it through. This restores baseline-matching resolution and recovers test_final_states / test_lin_probs (which now pass). The new param is backward-compatible — all existing callers behave identically.

Stale baselines (7 xfail)

Removing the resize surfaced 7 committed baselines it had been masking. They are genuinely stale and need regenerating on Linux CI (they can't be reliably regenerated on macOS — font metrics shift the tight bbox). They're marked xfail(strict=False) with reasons so they're tracked, not masked, and pass without XPASS errors once regenerated:

• 5 GPCCA macrostate-scatter (test_gpcca_meta_states*, test_gpcca_final_states) — scvelo→scanpy content drift (Replace scv.pl.scatter and scvelo paga with scanpy equivalents #1302, ~69 RMS). The size half is now fixed by the dpi change; the content half remains.
• test_proj_default_ordering, test_msc_default — later layout/figsize changes.

Coordination with #1328

These 7 should be regenerated and un-xfailed as part of #1328 (they are the canonical visual-regression tests #1328 keeps, not knob tests it converts). Ordering matters: the dpi fix must be in the tree before those macrostate-scatter baselines are regenerated, otherwise the oversized renders get baked in and have to be regenerated twice. strict=False means there's no race if a regenerated baseline lands before its marker is removed.

Testing

Local (macOS): 298 passed, 7 xfailed, 0 failed for tests/test_plotting.py; tests/test_gpcca.py plotting paths 32 passed. pre-commit clean.

🤖 Generated with Claude Code