Expose a /metrics endpoint on the proxy that returns cost, token count, and request volume data in Prometheus exposition format. Teams already running Grafana or Datadog shouldn't need the dashboard to monitor LLMKit - they should be able to scrape the proxy directly. Scope it to per-key and per-model breakdowns.