[szibis]

Summary

Stacked on #796 (daily-PR skip). Merge #796 first; this PR then retargets to main.

The daily performance run now reports what changed since the previous run, and turns red on a regression:

• Regressions:
  • a result that goes from same as Loki to different;
  • a status that goes from 200 to an error;
  • a slowdown that holds two runs in a row, or that the same run's release comparison also shows.
• Improvements: the reverse moves.
• Ratchet losses: a row that was at least as fast as Loki (Loki's cold computed, not cache-served) and had Loki's result, and lost that.

Each alert names the commit range to bisect. The alerts go at the top of the daily PR body and into the job summary. A final step fails the run when there are regressions or ratchet losses, so GitHub notifies the maintainer. No issues are opened and no labels are added.

Changes

• bench/ab/history.py:
  • a loki_cold_computed history column, derived from cold_cache_served and Loki's cache counters, so old lines read as unknown;
  • an alerts subcommand (pure function alerts()), plus ratchet.json (add-only, sorted).
  • Rules:
    • a slowdown counts only if it held two runs, or the same run's verdict vs the release is slower;
    • a cold ratchet loss needs two runs;
    • parity and status losses are immediate;
    • only sets the run wrote are evaluated, so a one-set manual dispatch does not re-report other sets;
    • warm p50 is never ratcheted, because Loki's warm answer comes from its results cache.
• bench/ab/daily.py: computes alerts after history is appended, updates ratchet.json, and writes alerts.md and alerts.json under --out.
• perf-daily.yaml:
  • the alerts go first in the PR body and in the job summary;
  • a Fail on regressions step runs after the PR and arm-auto-merge steps, with values passed through the environment;
  • alerts.md is uploaded with the run artifacts.
• bench/ab/README.md and a CHANGELOG entry under CI.

Proofs

• bench/ab/tests: 124 pass, including 24 alert tests. They cover each regression, improvement and ratchet rule, the two-run requirements, the cache-served and unknown Loki colds that never ratchet, old lines without the new column, the commit ranges, stable markdown, and the CLI writing its files.
• Script tests: scripts/ci/tests: 77 pass, including the pinned-uses test.
• Committed history: alerts reports no regressions and no improvements. The ratchet starts empty, because committed lines lack the new column; it fills from the first daily run after merge.
• trend.md, performance.md, gaps.md: regenerate with no diff.
• Linting and gates: actionlint (pinned and local) and zizmor 1.30.1 (pinned) report no findings; the conformance and changelog gates pass.

Not applicable: no proxy code changes, so e2e vs Loki, the Grafana range matrix, visual proof, A/B, benchstat and the load test do not apply.