usage_python vignette fails with mudata >= 0.4

[rcannood]

The usage_python vignette fails on CI since 2026-08-17, breaking R-CMD-check, r-universe and pkgdown on every branch:

Quitting from usage_python.Rmd:119-127 [load_mudata_example]
! anndata._io.specs.registry.IORegistryError: No read method registered for IOSpec(encoding_type='AnnData', encoding_version='0.1.0') from <class 'h5py._hl.group.Group'>. You may need to update your installation of anndata.

Cause

mudata 0.4.0 (released 2026-08-17) rewrote read_h5mu() to dispatch modality groups through anndata's IO registry on the encoding-type attribute. The example file used in the vignette (minipbcite.h5mu) was written by mudata <= 0.1.1 and stores encoding-type: "AnnData", while the registry only knows the lowercase "anndata" that mudata has written since 0.1.2. mudata's own backwards-compatibility test archives only go back to v0.1.2, so this went unnoticed upstream.

Verified locally:

• mudata 0.3.10 + anndata 0.12: reads the file fine
• mudata 0.4.0 / 0.4.1, with anndata 0.12 or 0.13: IORegistryError
• mudata 0.4.1 after rewriting the attribute to "anndata": reads the file fine

Minimal reproducible example (no R needed):

curl -sLO https://github.com/gtca/h5xx-datasets/raw/b1177ac8877c89d8bb355b072164384b4e9cc81d/datasets/minipbcite.h5mu
uv run --with "mudata==0.4.1" python -c "import mudata; mudata.read_h5mu('minipbcite.h5mu')"

Plan

• Pin mudata<0.4 in the vignette so CI is green again
• Report the regression upstream to scverse/mudata -- a fallback for the legacy capitalised encoding-type would fix it -- or switch the vignette to an example file written by a recent mudata
• Remove the pin once either of those lands