fix(lifecycle): gate Codex research on fresh evidence and preserve risk bans - #595

Old or undated drift observations could be treated as current and start research. The probe now preserves observation time and matching source identity, uses an explicit replay clock and a seven-calendar-day default validity window, and rejects stale, future or unavailable evidence. Existing REVIEW/CRITICAL risk prohibitions remain effective even when research is unavailable. Missing dates remain null across storage and reporting.

The autopilot binds to the existing actionable research cycle with subscription-based diagnosis, bounded search, strict WFA/OOS evidence, paired shadow and confirmed console delivery. Research defaults to Codex; AI_GATEWAY_RESEARCH_PROVIDERS explicitly permits Cursor or Codex→Cursor admission routing through the gateway. Missing callbacks stop before AI/optimization, quota deferral is retained, and human acceptance does not grant live authority. Old gateways/SDKs cannot silently accept the new route; admitted task and route identity are checked, failures are sanitized, and a submitted job never retries through the reliability/API fallback.

Console delivery now reads the exact ticket before and after creation. A confirmed 404 permits one POST; unavailable reads do not authorize writes, and an uncertain write is reconciled without another POST from the same callback. Full candidate, budget, evidence and state must match, including Python/JavaScript numeric serialization. The normal research cycle and CLI share saved-ticket decision recovery; matching accept/reject decisions are saved atomically as intent only, while pending or just-recovered profiles skip new research in that cycle. Existing terminal tickets are idempotent no-ops; cross-run experiment dispatch/deduplication remains separate integration work.

Validation: 219 tests and 14 subtests passed across 16 related files, including 32 freshness cases. Full Ruff passed after removing one duplicate test import; the affected 11 probe tests also passed. Independent review checked date/source consumers and ran 12 additional offline assertions. Existing strict cycle and paired-shadow fixtures passed; they are not real forward-shadow evidence.

Decision-recovery follow-up: 171 related tests and 97 subtests passed, including 46 new regression cases. Independent localhost HTTP integration used the real QRT Worker with synthetic in-memory KV/session to verify submit/readback, duplicate suppression, anonymous/stale-candidate rejection, accept/reject recovery and no live authority. Each case made exactly one ticket POST. This validates application wiring, not deployed authentication, durable cloud KV or real research/shadow evidence.

Subscription consumer validation: four related files passed with 103 tests and 26 subtests. Independent review reproduced incorrect task binding and raw error propagation; both now have failing-before/fixed-after regressions and independent closure. Ruff and diff checks passed. These are offline checks, not Cursor model or billing validation.

Adoption: deploy/install the service and SDK from QuantStrategyLab/AIAuditBridge#165 before enabling the consumer route. Adopt the new probe and risk mapper together from one QPK package; copying only the probe can lose retained risk prohibitions. No deployment, dependency-pin rollout, model call or trading action is performed by this PR. A real experiment job and approved data/shadow bindings remain outstanding.