observability-instrumentation

Wrap surfaces with Sentry.withErrorBoundary at the app shell, each route outlet, and lazy chunks, providing a fallback component and low-cardinality tags like surface and route. Initialize the SDK lazily via dynamic import so it stays out of the first-paint chunk.

Subscribe with onLCP, onCLS, and onINP from the web-vitals package inside an injectable WebVitalsReporter class, then forward each metric as a typed event through the reporter interface. Pair the signals with Lighthouse desktop and mobile audits so lab and field data agree.

Yes, keep the same event classes, boundaries, and reporter interface but register a structured-log reporter that writes one JSON object per signal, or a no-op reporter. Swapping to a live RUM sink later only changes the reporter implementation and its DI registration.

Never capture passwords, tokens, cookies, auth headers, raw form values, full API payloads, or user-entered text. Avoid high-cardinality tags like user IDs, session IDs, and full URLs with params, since they inflate RUM costs and risk leaking PII.

Register a reporter spy against the telemetry token in the tsyringe container, resolve the reporter or render the error boundary, and assert captured event names, values, and context precisely. Re-register the spy per test rather than clearing module-level singletons.

Do not duplicate infrastructure metrics like server latency, error rates, RPS, or backend business KPIs that the platform or APM agent already emits. Browser telemetry should cover only client-side failures and user-perceived performance signals.