The deterministic coordinator reserves at most three PR attempts and persists each intent before dispatch. It reconciles actual worker runs, keeps active locks through completion, and gives the oldest attempted PRs priority so the backlog makes progress.

Fresh current-head failure evidence updates the no-progress streak before another repair is selected. Dispatch errors retain durable intent rather than creating duplicate work. This layer remains an injected, testable policy module; Actions wiring follows later.

Validation: coordinator tests cover fairness, lost responses/results, identity mismatches, drift, capacity, and repeated published failures. Independently reviewed.