Clear code smell flagged by vulture, jscpd, and radon/xenon

[aarmey]

Ran the maintainability tools from CLAUDE.md across scrise/ and analysis/ and addressed what they flagged. (Note: CLAUDE.md refers to a cccrise/ package; the source actually lives in scrise/ and analysis/, and vulture/radon/xenon aren't in the dev group, so I ran them via uv run --with ... rather than touching pyproject.toml.)

Dead code — vulture

Most of the 37 hits were false positives (pytest fixtures, the fbuild entry point, figure helpers vulture can't trace across modules). Three were real:

• Dropped analysis.gating.getHiResOldLupus and the cell_type_conv table it alone consumed — no caller, and no dataset in the repo carries the "Cell Type Old" column it reads.
• Dropped figure3.plot_cell_perc_corr — no caller.
• gateThomsonCellsLeiden, Thomson_Doublet and the marker_genes_* tables are the hand-run cache-builder/reference-data exception CLAUDE.md describes. Left in place, with a short note saying why they have no in-repo caller.

Duplication — jscpd

13 clones (1.02% of Python lines), 0 in analysis/. Three were genuine copy-paste and got shared helpers:

• plotting/factors.py — _draw_unit_scaled_heatmap backs plot_eigenstate_factors and plot_gene_factors (16-line clone).
• plotting/pacmap.py — _plot_continuous_pacmap backs plot_gene_pacmap and plot_wp_pacmap (11-line clone).
• plotting/general.py — _two_gene_frame backs gene_plot_cells and plot_cell_gene_corr (10-line clone).

Three others were left alone on purpose: the _pf2_utils.py↔factorization.py hit is a pass-through kwarg signature, and the other two are repeated docstring parameter blocks. Deduplicating those costs documentation without saving logic.

Complexity — radon / xenon

Maintainability index was rank A for every file. Seven functions ranked C; the three in library code are now below threshold:

• rank_selection._resolve_bicv_inputs (C-14) → _resolve_dataset_alias, _validate_split_params, _ensure_condition_idxs.
• annotation_alignment.score_cell_type_alignment (C-11) → per-component loop extracted as _component_metrics.
• factor_io._attach_raw_data (C-11) → _subset_raw_cells / _subset_raw_genes.

Still rank C: three test functions and figureS9a_d.run_benchmarks. These are pre-existing, their complexity is assertion/parametrization chains where splitting hurts readability, and CLAUDE.md says to leave those.

Along the way

• Docstring drift: an AST check of public signatures against their Parameters blocks found three functions with undocumented parameters — pf2 and bicv (condition_key, adata) and plot_condition_factors (control_pattern, control_conditions). Now documented.
• Wrong annotations (ty): groupDrugs and plot_correlation_cmp_cell_count_perc annotated X as the anndata module rather than anndata.AnnData; figureS9e_f.plot_logistic_regression_component declared -> pd.DataFrame while returning a tuple. ty now reports zero diagnostics across scrise/.
• 24 unused unpacked fig variables in the plotting tests, flagged by both vulture and ruff's RUF059. This was the highest-value item: it was drowning vulture's output, which drops from 37 hits to 11 — all 11 now genuine documented exceptions, so the report is worth reading again. Also removed one stale # noqa: E402.

Verification

ruff check and ruff format clean; ty check scrise/ clean. Full suite: 133 passed, 1 skipped. Every refactored function has existing smoke-test coverage.

Deliberately out of scope

Flagging these rather than folding them in, since each is a judgment call:

• 45 NPY002 legacy np.random uses (mostly figure6.py and tests) — conftest.py already uses default_rng, so the codebase is split, but migrating changes the numbers drawn and would move the golden-regression baselines.
• camelCase in the public API (calculateFMS; params cellType, doEmbedding, cbarMax, labelType, ThomsonNorm) — renaming is breaking.
• The X/adata dual-alias params on pf2/bicv — part of why _resolve_bicv_inputs ranked C; worth retiring if one is a compatibility shim.
• factor_io.py:110 catches ImportError, AttributeError, KeyError, ValueError, OSError for the vsparse→read_h5ad fallback. Only ImportError is the intended case; the rest silently turn genuine data errors into a path change.
• 27 remaining ty diagnostics in analysis/ — 5 uninstalled optional deps (cupy, scib, scanorama, harmonypy), most of the rest anndata DataFrame | Dataset2D union noise.
• 16 RET504 assign-then-return, mostly in figure scripts — churn.
• Figure-script clones at lower jscpd thresholds — all the standard makeFigure preamble, i.e. the intended template.

🤖 Generated with Claude Code