Prometheus scraping, application metrics and Grafana dashboards

[FullFran]

prom-client is already a dependency, so metrics exist. What is missing is the
Kubernetes-side plumbing and something worth looking at.

Scope

• Confirm every runtime exposes /metrics, including worker and scheduler
• Install kube-prometheus-stack
• ServiceMonitor (or PodMonitor for the non-HTTP runtimes) per runtime
• Application metrics that actually matter, not just process defaults:
  queue depth, job duration histogram, job failure rate, feed fetch latency,
  alerts delivered
• Grafana dashboard: queue depth, worker replicas, job throughput, error rate,
  end-to-end alert latency
• Alert rules: backlog growing for N minutes, failure rate above threshold,
  scheduler missed a run

Notes

The dashboard should make the KEDA autoscaling issue visible — replicas and
queue depth on the same time axis is the single most convincing screenshot this
project can produce.