feat(runtime): evaluate Memory write-gate promotion evidence #1930

Scope

Track the concrete promotion evidence and offline evaluation artifacts required before the RFC #1770 Memory write evidence-sufficiency policy moves from shadow or advisory to enforcing.

This is follow-up work to RFC #1770 and PR #1921. It does not require replay as a runtime feature for a consumer that remains in shadow or advisory, and it is not a commitment to promote the policy, enable a hosted backend, or add any other decision-policy consumer.

Problem

PR #1921 defines the Memory write policy, defaults new policies to shadow, and persists scope-private observation sidecars with structured references and bounded policy/action/fallback/privacy facts. That establishes an observation baseline without retaining raw candidate or evidence bodies.

The policy's declared promotion_criteria are not measured promotion evidence. The current implementation does not provide a reproducible replay/rescore or re-evaluation workflow, a labeled calibration set, separate held-out evaluation, or a documented promotion and rollback decision. Those missing artifacts mean the observations may inform a future decision but cannot justify enforcing; enabling it merely because configuration permits the mode would treat configuration as evidence.

Acceptance criteria

• Define a reproducible, authorization-aware replay or re-evaluation workflow for the Memory write gate, reporting incomplete input coverage separately from negative results.
• Provide a labeled calibration set and separate held-out evaluation, including English and Chinese cases where the gate accepts both.
• Evaluate candidate outcome mappings and confidence thresholds against false holds, false passes, fallback reasons, latency, request cost, and hold/review friction.
• Document the measured promotion decision and rollback procedure, or explicitly retain shadow/advisory when the evidence is insufficient.
• Preserve immutable Memory revisions, exact citations, Scope isolation, and the invariant used_fallback=true => coverage=unadjudicated.

Non-goals

• No automatic promotion to enforcing.
• No dashboard, public HTTP/OpenAPI surface, export API, generic decision workflow engine, hosted-redacted transport, query rewrite, or Handoff/Experience/Skill/duplicate-conflict consumer.
• No retention of raw candidate text, evidence text, model rationale, or exception details for convenience.

References

• RFC docs(rfc): propose decision policy governance #1770
• PR feat(runtime): add decision policy governance types #1921