ci: fix red CI on main (zarr pin, heatmap baseline, scanpy 1.13)

[Marius1311]

Problem

CI has been red on main since the start of August, blocking the All tests pass gate on every PR. Three independent causes:

1. PETSc / SLEPc + R (conda job) — fails at conftest import:

   ModuleNotFoundError: No module named 'zarr.core.dtype'; 'zarr.core' is not a package
   

   The job installs cellrank (pip install -e '.[plot]', pulling anndata 0.13.2, which requires zarr>=3.1), then the next line force-downgrades to 'zarr<3'. anndata 0.13 imports zarr.core.dtype, a zarr-v3-only module, so the downgrade breaks the whole job. The pin dates to CI workflow modernization #1291 (Feb 2026), when anndata still supported zarr v2 — it's now stale.

2. All hatch-test stable legs (Linux 3.12/3.14 and macOS) — TestHeatmap::test_heatmap_genes fails with RMS 67.4 against a 50 tolerance, identically on all three. The baseline was rendered with matplotlib 3.11.0; 3.11.1 (in the env since the 2026-08-01 scheduled run) shifts the constrained_layout placement of the genes-mode heatmap. The plotted content is unchanged — only the layout moved.

3. hatch-test.py3.14-pre — collection of the entire suite fails:

   ModuleNotFoundError: No module named 'scanpy.plotting.palettes'
   

   scanpy 1.13.0a1 moves the legacy plotting API under scanpy.plotting.legacy. cellrank reached into two private scanpy symbols: palettes.vega_20_scanpy and _utils.add_colors_for_categorical_sample_annotation.

Changes

• Drop the zarr<3 pin in the conda job and let anndata resolve zarr (>=3.1).
• Refresh the heatmap_genes baseline, promoted from the Linux py3.12-stable rendered-figures artifact as docs/contributing.md prescribes. Reviewed against the old baseline: same curves, same ordering, same colour mapping — layout offset only.
• Stop importing scanpy's private plotting internals rather than chasing their new location. The palette is a fixed reordering of matplotlib's tab20, so it now lives in _utils/_colors.py, pinned by a test (the committed figure baselines depend on those exact values, and it is byte-identical to scanpy's vega_20_scanpy in both 1.12 and 1.13). The colour assignment is what cellrank's own _get_categorical_colors already did for the plotting functions, so that helper moves from pl/_utils to _utils/_colors — models/ can't import from pl/ without a cycle — and gains the one behaviour scanpy's helper contributed: regenerating palettes that are too short or not colour-like.
• Keep the macOS leg blocking. The pyGPCCA brandts failures that originally motivated continue-on-error for macOS no longer reproduce — on the latest main run macOS is green apart from the shared baseline drift, so it stays a gating job.

Verification

Locally on Linux, both matrix envs pass the full suite: py3.12-stable (matplotlib 3.11.1, scanpy 1.12.3) and py3.14-pre (scanpy 1.13.0a1), the latter having previously failed at collection. The refreshed baseline is byte-identical to what this machine renders.

🤖 Generated with Claude Code