Influence diagnostics: dffits + influence masks (#27 batch 3)

[sipemu]

Batch 3 of 4 for #27. Adds 4 influence diagnostics as Polars expressions.

Implemented

Diagnostic	Expression	Default threshold	Output shape
DFFITS	`ps.dffits("y", "x1", ...)`	—	struct{dffits: List, n_observations}
Influential Cook's	`ps.influential_cooks("y", "x1", ...)`	4/n	struct{is_influential: List, n_influential, n_observations}
Influential DFFITS	`ps.influential_dffits("y", "x1", ...)`	2 * √(p/n)	same
High leverage points	`ps.high_leverage_points("x1", "x2", ...)`	2 * p/n	same

All four follow the rlib split pattern (public *_fit + pl_* shim). Each fits OLS internally via the shared fit_ols_for_residual_diag helper introduced in batch 1, except high_leverage_points which only builds the design matrix (no y needed).

Tests

• Python — tests/test_influence_diagnostics.py, 7 new pytests:
  • DFFITS at planted high-influence indices > 1.0.
  • influential_cooks and influential_dffits flag the planted indices.
  • Custom threshold tightens the influence mask.
  • high_leverage_points catches extreme x values without taking y.
• Rust — tests/rust_api.rs: smoke tests in diagnostic_fits for all 4 *_fit entry points.

Full Python suite: 448/448 pass (441 prior + 7 new). Rust integration: 12/12 pass.

Remaining work on #27

• Batch 4 — multicollinearity + dispersion: generalized_vif, high_vif_predictors, classify_condition_number, variance_decomposition_proportions, pearson_chi_squared, estimate_dispersion_pearson, estimate_dispersion_deviance.
• Probit/cloglog/negative_binomial/Tweedie residual variants (deferred from batch 2).
• Standardized Pearson / standardized deviance residuals.
• compute_leverage_with_aliased (variant for rank-deficient designs).

Test plan

• cargo build --no-default-features clean
• cargo clippy --all-targets -- -D warnings clean
• cargo fmt --check clean
• cargo test --no-default-features — 12/12 pass
• pytest — 448/448 pass
• CI matrix

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!