RabbitMQ redelivery storm from timeouts

RabbitMQ cluster, quorum queue image-resize. At 11:30 the redelivered message rate on the queue spikes to ~95% of all deliveries, the queue's ready count oscillates between 40k and 80k, and consumer unacked counts on each of the 8 workers sit pinned at their prefetch limit (50). Throughput (acks/sec) has fallen ~70%. No deploy. Recent context: the upstream object store the resize workers read from started returning intermittent 5s timeouts about 15 minutes before the alert; the broker's per-message delivery-attempts/x-death header counts are climbing. How do you triage and mitigate this redelivery storm?