Run T+24h and T+7d re-checks on every eligible deployment_verification finding.

Called by an EventBridge hourly rule (see scripts/deploy/create_verify_deployment_schedule.sh). Deterministic, no LLM, no agent session — just: find eligible findings, resolve the target metric, PATCH the observation, file an anomaly finding if the observation regressed past the monitoring plan's threshold.

The returned summary lets EventBridge logs show the cycle's work and gives Platform EM's next session a quick signal via the findings stream (regression anomalies appear in fetch_unread_signals_for_me).

Safe to call ad hoc for debugging — idempotent (already-observed findings are excluded by the eligibility query).