Versioned evaluation workflows need durable lineage and replay-equivalent failures

Reusable implementation insight (observed 2026-09-05, source: COR-4055 implementation and independent review in ender-inference): treat reevaluation and human review as append-only state transitions, not compatibility-side effects. A reevaluation must persist its child run before updating any mutable latest-grade projection; failed idempotent replays must reproduce the original failure semantics/status; and concurrent reviews need a serialized per-finding append point with the same deterministic (created_at, review_id) tie-break in API and database code. Review corrections should normalize result/applicability as one consistent pair. This combination prevents false success, branching review histories, and replay behavior that changes across retries.