HPA scales on CPU but the true bottleneck is a single-threaded queue consumer partition assignment

Autoscaling appears to help, yet throughput never rises because the workload's shard or partition ownership keeps all useful work pinned to one replica.