performance-engineeringlisted

# Performance Engineering The performance & scale conscience of the build: prove the system meets its latency, throughput, error-rate, and capacity targets — with numbers, not vibes. Honors `${CLAUDE_PLUGIN_ROOT}/skills/_shared/vibegod-principles.md`; priority is **user > skills > default**. This skill proves performance; it does NOT own deploy infra (that's `devops-sre`). ## Fits in the pipeline - **Stage 3 (Stack & cost)** — supply the performance NFRs the architecture must hit (SLO targets, expected/peak/spike traffic), so the stack is chosen against real load, not guesses. - **Stage 7 (Per-feature QA gate)** — the performance lens: run the relevant profile(s) against the feature and fail the gate on any breached threshold or regression vs baseline. - **Stage 8 (Ship gate)** — the binding pre-ship proof: full load/stress/soak/spike sign-off, breaking-point + capacity headroom, and p75 field Core Web Vitals for user-facing surfaces. A "performance tested" claim is only valid here when it is reproducible, thresholded, and run on a representative system. ## Best practices - **Gate on percentiles, never averages.** Every latency SLO is p95 and/or p99 (p99.9 for critical paths). Averages are forbidden as pass/fail criteria. - **Codify SLOs as machine-checkable thresholds.** Each test declares them, e.g. `http_req_duration: ['p(95)<200','p(99)<300']`, `http_req_failed: ['rate<0.01']`, `checks: ['rate>0.99']`. A breached threshold MUST produce a non-zero exit so CI