feat(observability): add metrics, SLOs, job health, and provider telemetry #698

Parent epic: #679
Related logging foundation: #447

Objective

Add provider-neutral, privacy-safe operational signals for the web runtime, database pool, background jobs, and optional external providers.

Scope

• Define a small metrics port with no-op/default implementation.
• Add low-cardinality counters/histograms/gauges for request outcomes/latency, pool saturation, job run status/backlog, and provider outcome/latency/circuit state.
• Define initial SLIs/SLOs and burn-rate/threshold guidance.
• Add authorized health/readiness output that distinguishes local dependencies from optional online providers.
• Document Prometheus/OpenTelemetry adapter examples as optional integrations.
• Add dashboard/runbook examples without coupling core runtime to a SaaS.

Guardrails

• No tenant IDs, routes with unbounded IDs, email/IP, object keys, tokens, prompts, or conversation content in labels.
• Offline/LAN operation works with no external collector.
• Metrics must not be used as an authorization source.

Acceptance criteria

• Metric names/labels have documented cardinality and privacy review.
• Job/provider failure and recovery are observable.
• SLOs map to actionable runbooks.
• Unit tests cover no-op and representative adapter behavior.
• Load/smoke test proves metrics do not materially degrade request processing.