list_alerts ordered by created_at DESC alone; kpi_alerts has many rows sharing the same second timestamp (07:00 batches), so MySQL tie order is unstable and offset pagination both duplicated and skipped rows (entity 1 read-only probe: total=281, 281 rows walked, only 268 unique ids, 13 duplicates; id 587 unreachable in the unfiltered list while visible in ?status=pending).
Add secondary key id DESC so pages form a total order; add regression suite test_alerts_pagination_stability_20260925.py (tie-heavy fixture: unique==total, zero duplicates, strict id DESC inside a tie group, page concat == single-shot ordered query, unfiltered vs status-filtered visibility, entity isolation negative control).
GET /api/cma/alerts/check-timeout picked candidates via the join on KPIDefinition.entity_id, so a session could batch-upgrade (red + processing) alert rows whose own entity_id belongs to another account. Switch the filter to the alert row's own entity_id, drop the now-unneeded join and the unused KPIDefinition import. Write path only; no data change, candidate ids on consistent rows are identical before/after.
Derived from dama-20260919-cma-alert-filter-and-cash-baseline-fix (C1/C2 residual in the write endpoint).
alert-to-plan, deviation-push review topic and AI-suggestion-to-plan built ActionPlan(...) without entity_id, so non-e1 sessions wrote rows with the model default 1 and entity_id stayed untrustworthy on those paths. Write the session entity at all three sites, scope the deviation-push topic dedupe by entity (a foreign same-title topic no longer suppresses the session entity's own topic), and require the applied suggestion to belong to the session entity. Adds 7 regression tests (both directions + cross-entity negative controls + auth floor).