Two details here deserve more attention than they usually get: separating successful and failed latency, since fast errors can make the success picture look healthier than it is, and checking the observer itself - quiet traffic and missing visibility look identical until telemetry freshness or a synthetic check says otherwise.