RFAI-05: Confidence pipeline and Review evidence section

[Chris0Jeky]

Context

Week 5 from taskdeck-12-week-roadmap-v4.md.

Parent: #972
Depends on: #976

This issue makes proposal confidence and evidence visible enough for review-first automation to feel trustworthy, not opaque.

Scope

• Combine verbalized confidence, provider confidence/logprob signals where available, and attribution/provenance verification scores into field-level confidence.
• Add Review UI evidence sections using source spans and EvidenceLink reason chips.
• Add self-consistency only for high-criticality low-confidence decisions, with explicit quotas/cost guardrails.
• Add Brier-score measurement on confidence buckets for the golden set.
• Preserve existing readable proposal cards, sticky actions, and board-centered review flow; this is an evidence depth pass, not another Review decomposition.

Acceptance Criteria

• Review cards expose why a proposal exists, with source/evidence links where available.
• Confidence buckets are generated deterministically and can be scored against the golden set.
• High-criticality low-confidence self-consistency is gated by quotas and never silently applies work.
• Manual inbox-triage review evidence path is documented and can be tested.
• Brier score target and current result are reported by the eval harness.

Suggested Verification

• Frontend Review component tests
• Backend confidence aggregation tests
• Golden eval smoke including confidence-bucket scoring
• Manual Review flow check recorded in docs when behavior lands