feat(knee-point): noise-estimated ε helper

[jramos]

Summary

Adds a pure-function helper _estimate_val_noise(...) to evolution/skills/knee_point.py that estimates a noise-floor ε for the Pareto knee-point band via a paired bootstrap on per-example per-candidate val scores (DSPy's DspyGEPAResult.val_subscores).

This PR introduces no behavior change. The helper is unused in production code — confirmed by grep -rn '_estimate_val_noise' evolution/ returning only the definition. A follow-up PR will plumb it through select_knee_point behind a --knee-point-epsilon noise-estimated sentinel, gated on empirical evidence from a regenerated calibration campaign.

Motivation

The current knee-point selector uses a fixed ε = 1.0 / n_val band. Investigation showed this is statistically incoherent:

• At our typical n_val (8–50), the actual paired-bootstrap noise floor on val scores is ~0.04 empirically.
• 1/n_val is ~0.02–0.125 — an order of magnitude below the real noise.
• The May 2026 calibration data shows our knee-point selector picks the same candidate as GEPA's val-argmax in 7 of 9 runs at the current ε — because the band rarely separates from val-argmax when ε is sub-noise.

The GP-parsimony literature (Poli & McPhee, "Parsimony Pressure Made Easy," GECCO 2008) treats the ε-band as a hyperparameter that should be set from a noise estimate. The new helper computes that estimate from data DSPy already returns.

Helper contract

• Single candidate: returns 0.5 / sqrt(n_val) (worst-case binomial SE at p=0.5) as a degenerate-path floor.
• All-saturated scores: returns 0.0 (no useful signal → no band).
• Otherwise: paired bootstrap on the concatenated pairwise diff vector between best and each competitor; returns half the confidence-level CI width on the mean (default 90% CI, 1000 resamples, seeded for determinism).
• Coverage policy: positional alignment with min(len(best), len(other)) truncation.

Stdlib only (math, random). No numpy/scipy.

Test plan

• 6 unit tests in tests/skills/test_knee_point_noise_estimation.py: saturated → 0, Bernoulli analytical-SE reference (order-of-magnitude pin at p=0.5, n=50; paired-diff SE not single-Bernoulli SE), monotone-in-variance, single-candidate fallback (exact value at n=64), determinism, partial-coverage non-crash.
• Full suite green locally (1150 passed).
• CI green across all Python versions.

What's next

A regenerated calibration campaign (~$140, out-of-band) will produce gate_decision.json + band_holdout.json artifacts. An extended study_b_pick_epsilon.py will replay the selector with three modes — current 1/n_val, the original calibration recommendation 0.5/n_val, and this PR's noise-estimated mode — to decide empirically whether the noise-estimated default beats the geometric one on real per-skill holdout pick quality. If yes, a follow-up PR plumbs the sentinel through. If no, the right ship is to drop the selector for val-best and defer to GEPA's best_idx.