Resume a failed job from its last durable boundary.

State-aware and replay-safe. Active jobs—including RESERVED jobs in centralized provider backoff—are returned unchanged, except a RUNNING table read with runStale: true, whose run is sent again from its last checkpoint at no new cost (generation_requeued). Terminal legacy/pre-worker provider failures may reserve their original quote and re-enter generation under the same job id. READY/PARTIAL jobs with a failed table-read finalize retry that existing artifact without regenerating it. Other started generation failures are not blindly restarted because they may own durable intermediate resources; canceled jobs cannot be resumed. A resume that would reserve or re-enter generation asks the project's gate again first: when it has closed since the job was created, the call is refused with 409 project_precondition_failed and nothing is reserved. A running job asks it again before each paid step too, and a gate that closed mid-run fails the job with failureCode: project_precondition_failed, releasing its credits. Requires Idempotency-Key.