feat(stats): reviewer-proof correlation helpers + prisma correlate

[javiercuervo]

What

Adds a new prisma.stats subpackage and a prisma correlate CLI command: correlation and robustness statistics designed so every estimate ships with what a sceptical reviewer asks for.

• fisher_ci(r, n) — Fisher z-transform 95% CI for a Pearson r.
• correlation_table(df, x, y, group=None) — Pearson r + Fisher-z CI and Spearman ρ, with n, overall and per stratum (the Pearson/Spearman gap flags outlier-driven associations).
• partial_correlation(df, x, y, covar) — first-order partial r(x,y|covar) (controls for confounders such as general ability).
• bootstrap_ci(df, x, y) — percentile bootstrap CI for small/fragile cells.
• missingness_compare(df, indicator, by) — selection-bias check (Mann-Whitney) between present vs missing rows.
• mixed_model_icc(df, outcome, predictor, group) — random-intercept model + ICC for clustered data (students nested in cohorts/courses); optional statsmodels.

Why

Extracted while hardening the statistical rigour of the 20-60-20 AI-permitted-assessment study (Cuervo, 2026, Universidade de Aveiro) against an adversarial three-reviewer panel: within-stratum reporting, rank robustness, partial correlation for mechanical/ability overlap, bootstrap for small cells, missingness and clustering. Reusable for any education/SLR dataset.

Notes

• Adds scipy as a core dependency; statsmodels is an optional [stats] extra (only mixed_model_icc needs it).
• New CLI: prisma correlate --in data.csv --x A --y B [--group G] [--partial Z] [--bootstrap] [--out table.csv].
• Tests: tests/test_stats.py (5 cases); full suite passes (16/16). CHANGELOG updated.

🤖 Generated with Claude Code

[javiercuervo]

CI arreglada: el step ruff check fallaba en src/prisma/stats/correlations.py con 3 errores que tumbaban los jobs 3.10/3.11/3.12. Aplicado en 3dae65d (un PR = una razón, solo lint): (1) F401 — eliminado Iterable, que estaba importado y sin usar; (2) UP035 — Sequence ahora se importa de collections.abc en lugar de typing (convención Python 3.9+); (3) RUF002 — sustituido el EN DASH ambiguo – por - en el docstring de partial_correlation ("x-y association"). ruff check pasa limpio en local. Al re-correr la CI debería ponerse verde. Nota: no toqué el WIP sin commitear (README.md, RELEASING.md, docs/citing-articles.md) — queda pendiente para su propio PR.