New internal/metrics package: AI metrics (summarize/caption/chat latency, LLM tokens), session metrics (requests+latency by bounded route pattern, logins), and a /metrics Handler. Adapters call a typed API; never touch prometheus types. Dep: github.com/prometheus/client_golang (standard Go client; cluster runs prometheus-operator). TDD: collectors + middleware route-pattern cardinality covered. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>