blog: how we stopped shipping performance regressions

[amavashev]

Summary

• New engineering blog post: How We Stopped Shipping Performance Regressions — the Cycles two-tier CI benchmark gate.
• Covers: release-blocking at 25% vs a pinned baseline; nightly trend-warning at 30% vs a rolling 7-run median; 3-trial medians to absorb GH-runner noise; [benchmark-skip] escape hatch; benchmark-data branch pattern; parse-layer failing loudly.
• Grounded in real source (parse-benchmarks.py, check-regression.py, release.yml, nightly-benchmark.yml, benchmarks/README.md) with verbatim code quotes.
• Dated 2026-04-18.

Review cycle

• Cycle 1: link verification, fact-check, SEO review in parallel. Broken forward-ref to unmerged TTL sweeper post stripped; p99 overclaim for Release/Event corrected; description trimmed for SEO.
• Cycle 2+3: self-review for flow; rated 9/10 — cited, specific, no hype.
• External reviewer: rated publishable. Five line-edits applied (tightened opener, softened empirical variance claim, softened operation-history claim, removed mild pattern-stealing overclaim, corrected "floor" to "bound").

Test plan

• npm run build succeeds (no dead-link errors)
• Post renders correctly in dev preview
• Internal cross-links resolve (benchmark post, Lua/Redis post, property tests post)
• GitHub source-code links resolve to main branch