Dead-letter queue replay

Design dead-letter and replay tooling for a large event platform. When a consumer fails a message N times it goes to a DLQ. You need operators to: inspect DLQ messages, see the failure reason and original headers, edit/fix a payload, and replay one message, a filtered subset, or an entire DLQ back into the original topic, without re-triggering already-succeeded downstream side effects or causing an infinite poison loop. Scale: 10K DLQ messages/min during a bad deploy, replay batches of millions.