Performance Regression Gate

The scheduled nightly-perf job protects the published 100k performance thresholds and absolute product SLOs. It deliberately keeps the existing pack.e2e_p95_ms < 300 contract instead of widening the limit when a shared GitHub-hosted runner produces an isolated pause.

Sampling and diagnostics

• The pack lane measures 50 indexed calls after warmup. This is large enough for one isolated host pause to remain visible as max_ms without dominating the p95 statistic.
• The receipt records the five slowest timed samples.
• The three slowest queries are replayed with pack --trace; replay totals and stage timings are diagnostic evidence and do not replace the original timed samples.

Confirmation policy

Scheduled and manually dispatched workflows run one confirmation attempt only when the first 100k gate fails. The job fails when the confirmation attempt also fails. Both attempt receipts are uploaded when confirmation is needed, preserving evidence for real regressions while reducing false alarms from a single hosted-runner stall.

Changelog

• 2026-07-27: increased pack sampling from 20 to 50, added slow-sample and traced-replay diagnostics, added one fail-confirmation attempt, and exposed manual workflow dispatch without changing stored thresholds or the 300 ms pack SLO.