[parashardhapola]

Exports could not carry analysis results, cluster labels exported as integers (so scanpy's rank_genes_groups failed), categoricals and colours were lost on import, and external latents could not be imported.

Changes

• to_anndata and to_h5ad take obsm, obs, graph, cell_selection, and dtype: PCA (with varm["PCs"] and uns["pca"]), UMAP, labels as categoricals with colours, and graphs in scanpy's obsp layout with uns["neighbors"], checked against the memory budget.
• Run exports write labels as string categoricals with the run's colours and add the run's PCA and Harmony correction.
• Imports keep categories, order, ordered, and colours; reduction.import_coordinates imports an external latent space, such as scVI's, aligned by cell ID.
• An index that is neither the ID nor the name column no longer imports as _index; count matrices are chosen by conventional name, with a warning when several hold counts; repeated obs names import with duplicate_cell_ids="make_unique"; to_mtx is faster.

Behaviour changes

• Exports of uint8 and uint16 counts are float32 unless dtype says otherwise.
• Run exports change shape (categorical labels, PCA added).
• H5AD inspection is stricter: negative integers are not counts, and a repeated index raises at import unless made unique.

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!