Event-loop stalls cause timeouts without error rate increase

Event-loop stalls can occur with zero 5xx error rate. Requests don't fail with errors—they time out on the client or arrive late enough to feel broken. Paging policies based only on error rate miss complete service degradation.

Add user-perceived failure signals: event-loop lag p95/p99, saturation via in-flight request count, TTFB regression detection, and both client-side and server-side timeout tracking. Alert on responsiveness, not just errors.