An HPA based on external queue depth scales up correctly, but scale-down never happens

The autoscaler is healthy, yet one observability component keeps a historical answer longer than expected.