bench: harden promotion with holdout uncertainty and stability evidence

[Lucas-Bur]

Parent: #166

Make candidate promotion depend on excluded evidence rather than fit-all diagnostics.

Acceptance criteria:

• Promotion status is derived from grouped and leave-one-repository-out holdouts plus an untouched final test or implemented nested cross-validation.
• Every rejected candidate reports exact guardrail blockers with partition, metric, candidate value, baseline, tolerance, and delta.
• Add grouped bootstrap or equivalent uncertainty estimates.
• Add local perturbation, plateau width, worst-case drop, and selection-frequency diagnostics across folds, seeds, and restarts.
• Preserve explicit no-eligible-candidate outcomes.

This issue remains benchmark-only and must not change Production retrieval behavior.

[Lucas-Bur]

Implemented and merged via PR #179.

Summary:

• Promotion is derived from grouped/LORO holdouts plus an untouched final fold; fit-all remains diagnostic.
• Rejected candidates retain exact blockers, bootstrap uncertainty, and stability evidence.
• Explicit no-eligible-candidate outcomes are preserved.
• Production retrieval behavior is unchanged.
• Merge commit: 24f728c

Validation:

• vp check: passed
• vp test: 467 passed, 1 skipped
• vp run lint:fallow: completed; only inherited repository health/dependency findings remain, no duplication
• vp run lint:fallow:ci: no new dead-code or duplication regression