[codex] add uncertainty-aware paired fit evidence

[divo12]

Stack

Depends on #14. Review against codex/confirmed-failure-curation.

Research finding

Agent outcomes vary across trials. Anthropic recommends repeated trials and separates pass@k from reliability; model-selection literature recommends uncertainty and multiple-comparison correction. OFW already paired champion and candidate by case/repeat, but it promoted from point estimates alone.

Implementation

• Adds typed PairedEvidence per partition: real wins, losses, ties, discordant pairs, net pass delta, candidate win rate, and exact one-sided sign-test probability.
• Excludes synthetic benchmark copies from significance claims while retaining their declared weight in effect-size metrics.
• Adds explicit StatisticalGateMode: EFFECT_SIZE_ONLY or EXACT_SIGN_TEST.
• Adds a frozen PairedEvidencePolicy to FitPolicy and all policy/cache digests.
• Rejects raw string modes at the direct Python runtime boundary instead of relying on annotations.
• Exact mode requires a declared minimum discordant-pair count and maximum probability.
• Applies Bonferroni correction across all candidates evaluated in the campaign.
• Keeps critical regression, regression score, target effect, latency, cost, selection, and admission gates unchanged and primary.
• Persists paired evidence in every CandidateOutcome, including rejected candidates.

TDD evidence

• Five real target wins produce exact probability 0.03125.
• Ties do not count as discordant evidence.
• Synthetic wins are excluded.
• Insufficient evidence fails exact mode.
• The same evidence fails after correction for two candidate comparisons at alpha 0.05.
• Five repeated target wins pass an end-to-end exact gate and reach one-shot admission.
• Effect-size-only mode preserves existing small-fixture behavior explicitly.

Verification

• pytest -q — 159 passed
• Ruff and strict mypy across src and tests — clean
• forbidden construct scan — no Any, raw dict, object, getattr, setattr, lambdas, or __import__
• Bandit — zero findings
• source distribution and wheel build successfully

Deliberate scope

• This is paired offline evidence, not a claim of randomized production A/B impact.
• No SciPy dependency, bootstrap framework, or online experimentation service is added.