Threshold regressions on PRs are human-overridable by design, so they should never generate task-failure telemetry. Hook/workload failures should generate alerts since nobody can rule those failures as acceptable. Today the hook marks the span as failed at threshold-detection time, before checking the PR for an authorised override, so accepted regressions still produce internal alerts.

Outcomes

• Overriding a failed threshold check produces no task-failure alerts and the task passes
• An un-overridden threshold failure still fails the task and still emits telemetry, just under a non-paging category
• Hook/workload failures continue to alert as today