gate: per-decision confidence signal via rotated-exemplar self-consistency

[JSONbored]

Context

Parent: #8828 (Phase 3 — mechanism). Calibrated abstention needs a real confidence score per decision. Verbalized confidence alone is poorly calibrated; sampling-based consistency works, and two samples capture most of the benefit; simulated annotators (same judge, rotated few-shot exemplars, inter-run agreement) outperformed both in the risk-control literature this epic builds on (Trust or Escalate, ICLR 2025).

Requirements

• For each AI-review gate judgment, run N=2-3 evaluations with rotated few-shot exemplars drawn from the golden corpus (sub-issue: golden replay corpus); the confidence score is inter-run agreement combined with verbalized confidence.
• Disagreement across runs → hold, surfaced with the mode-aware inconclusive copy (fix(ai-review): inconclusive-finding copy claims 'dual-model' in single+fallback mode, and claude_code auth errors lack structural circuit-breaker classification #8791 conventions).
• Persist the score with the decision (schema change: contiguous migration) so it joins the calibration set; export a bounded, anonymized form in the fleet signal only if privacy review passes — otherwise keep it instance-local.
• Cost control: the extra runs ride the existing budget accounting; a budget-exhausted state degrades to single-run + lower recorded confidence, never to a fabricated score.

Test Coverage Requirements

99%+ patch coverage, branch-counted; agreement scoring is pure and property-tested; budget-degradation arm covered.

Expected Outcome

Every decision carries a confidence signal with known empirical calibration — the input the abstention threshold needs.

[linear-code]

JSO-6112