feat(metrics): stdlib rank-correlation primitives (Spearman, Kendall, bootstrap CI)

[Denis-hamon]

Tier 2.1.a — the no-GPU foundation for the offline→downstream keystone (does a cheap planner-free offline metric predict control usefulness?). Pure stdlib, no compute, core wmel.metrics only.

What it adds

• spearman_rho — Pearson on average ranks (tie-safe).
• kendall_tau — tau-b, tie-corrected, O(n²); robust at very small n.
• bootstrap_correlation_ci → CorrelationResult — paired percentile bootstrap (resamples cell pairs jointly, like paired_bootstrap_gap_ci; skips degenerate resamples; deterministic given seed). Rank-based by design: a handful of cells, monotone-not-linear, incomparable scales.
• Exported from the package; _rankdata/_pearson private helpers.

Why now

T2.2 (n=150) and T2.3 (pixels) are GPU-gated, but this primitive layer is the planned no-GPU first step of the keystone and is independently useful stats. The remaining keystone pieces — offline_metrics.py (M1 1-step MSE, M2 k-step divergence, M3 action-counterfactual sensitivity, M4 score-to-go error) and correlate.py — need the CPU MLP-variant sweep + the GPU box for the TD-MPC2 cells.

Verification

• 13 tests, all closed-form: Spearman 0.8 and Kendall 4/6 on [1,2,3,4] vs [1,3,2,4]; tie-corrected tau-b 0.5; average-rank ties; perfect-correlation tight CI; determinism; degenerate + input-validation guards.
• Adversarial review: GO, zero must-fix — all three hand-derivable values re-derived and cross-checked against SciPy; paired resampling, percentile convention, determinism, and edge guards confirmed. Two doc nits applied (CI is conditional on non-degenerate resamples; n_boot field = valid count).
• Full suite passing; core stays stdlib-only; no emoji; no new git tag.

🤖 Generated with Claude Code