docs: document production monitoring, alerting, and degradation detection

[KostasCherv]

Gap

Observability docs covered AI tracing (LangSmith, LangFuse) well but said nothing about production monitoring. The Prometheus/Grafana stack is benchmark-only, and there was no documented metrics, alerting, uptime-check, or degradation-detection story for the Cloud Run deployment — nothing told an operator the service was degraded except a user report.

What this adds

New Production monitoring section in docs/observability.md:

• Cloud Run built-in metrics that apply (request count, latency percentiles, instance count, startup latency, CPU/memory) and where to view them.
• A recommended Cloud Monitoring alerting policy: 5xx ratio, p95 latency (with the SSE caveat), probe-failure log alert, and an uptime check against GET /health.
• "How degradation is detected": which signal fires for each failure class — dependency outage, elevated errors, latency regression, LLM-provider failure, ingestion/outbox stall, full outage.
• Notes the opt-in Sentry error capture (SENTRY_DSN) and clarifies the Prometheus/Grafana stack stays benchmark-only.

Also: README observability row updated to reference the documented policy instead of implying production dashboards, and a one-line cross-link from docs/deployment.md#health-and-readiness.

Note

Docs-only — this documents operational policy. Alert policies and uptime checks are per-project Cloud Monitoring resources, and the doc is explicit that they are recommended, not provisioned by the deploy scripts. No runtime code changes.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
cortex	Ready	Preview, Comment	Jul 14, 2026 2:26pm