ruff is clean at both scopes — and three of the reports were real

[shanikawm]

ruff check shanuz 51 → 0, ruff check . 201 → 0.

Most of it is what it looks like: unused imports left behind by the Assay5 refactor, dead locals, semicolon-joined statements. But three reports were pointing at something real, and those are the reason this is worth reading.

A test had been shadowed and never ran

test_anchors_seurat_parity.py defined test_pca_loadings_are_exact_not_randomized at line 275 and again at line 422 (F811). Python keeps the second, so the first was collected by nothing.

It is not a duplicate. The surviving copy checks _pca_loadings against np.linalg.svd and asserts orthonormality and seed-independence. The shadowed one additionally builds sklearn's randomized result and asserts it disagrees — the guard that the exact SVD is load-bearing rather than an equivalent spelling:

assert rand_pc.min() < 0.9, (
    "if randomized SVD agreed on every PC the exact solver would not matter"
)

That is the assertion behind the 45% → 100% RPCA anchor-recall fix, and it had not executed since it was written. Renamed to test_randomized_svd_would_visibly_disagree_on_the_trailing_pcs; it passes.

A regress-out test with no baseline

test_sctransform_vars_to_regress_removes_covariate_effect computed before (F841) and never asserted on it — and it was the wrong baseline regardless, taken from a different object built from un-injected counts. What remained was:

assert np.nanmean(corr_after) < 0.05

which is equally satisfied by a fixture that never carried the covariate in the first place. Rewritten to run SCTransform twice over the same injected counts, once plain and once regressing, and to require the signal to be present before requiring its removal: unregressed 0.2526, regressed 0.0000.

Verified the new half is load-bearing by zeroing the injection — corr_before drops to 0.073 and the test now fails there, where the old one passed.

find_markers named one missing ident but not the other

ident_1 empty said which ident; ident_2 empty raised a constant "No cells found for comparison group." (F541). Both name it now, and since ident_2=None means "every other ident", that case reports there is nothing left outside ident_1 instead of a parameter the caller never passed. Two new tests — the path had no coverage at all.

The rest

Dead locals worth naming: two computed-and-discarded colour maps in vln_plot and dim_plot, and a v3 layout flag in read_10x that no branch consumed.

The 121 E402s in tests/ and tutorials/ are all one deliberate line — the sys.path bootstrap those files need to run straight from a clone. Ignored by scope in pyproject.toml rather than by 121 scattered # noqa: E402.

Verification

	before	after
ruff check shanuz	51	0
ruff check .	201	0
suite	915 / 25 skipped	918 / 25 skipped
mypy shanuz	0	0

+1 from the un-shadowed test, +2 new. The pbmc3k tutorial was re-run end to end and its --report against Seurat 5.5.1 is unchanged (2,638 barcodes, 2,000 HVGs, kNN 52,760, 8 vs 9 clusters). 5 mutations, 5 killed.

One thing this does not fix

CI installs ruff unpinned and currently resolves 0.16.0, whose default rule set is much larger — uvx ruff@0.16.0 check shanuz reports 644, dominated by UP045/UP037/UP007 (Optional[X] → X | None, ~500 sites) and I001 import sorting. Everything above is measured against the pinned local 0.15.20. Closing that gap is a separate decision: pin ruff in CI, or take the annotation rewrite.

🤖 Generated with Claude Code