A dashboard can appear healthy because its own ingestion path stopped updating. I would show freshness timestamps and use an independent external check for the user-facing path. The monitor should make missing evidence visible rather than interpreting silence as success.

An average can look calm while a small set of requests is painfully slow. I would compare latency distributions by operation and time window, then inspect traces near the slow end. Low sample counts should be visible so the chart does not imply more certainty than exists.

An opaque identifier only helps if the client can retain it and the server can find it. My preferred error response carries the same request ID as the response header, with a short explanation and a stable error code. Logs should explain stages without retaining private request bodies.