perflisted

# Performance Work: $ARGUMENTS ## Principle **No optimization without a measurement — intuition names the wrong hotspot more often than not**, and a model's intuition is trained on other people's codebases. The profile decides where to work; the baseline decides whether it worked; the budget decides when to stop. An optimization that was never measured is a complexity purchase with an unknown price and an imagined benefit. ## Phase 0 — Load profile Read `.claude/PROJECT.md` (Commands — test/run/bench if present; Architecture; Integrations — external calls are frequent culprits) and applicable rules. Missing/TEMPLATE → fall back to the root `CLAUDE.md` (always in context) when it carries those facts, noting you're running without a kit profile; only if *neither* has them, STOP, `/bootstrap` first. ## Phase 1 — Budget and metric (before any code is read) - **One metric, named precisely**: p95 latency of endpoint X / wall-time of job Y / RSS after N ops / cold-start / throughput. "Faster" is not a metric. - **The budget**: the number at which this is *done* ("p95 < 300ms", "import of 100k rows < 2min"). From the user if it's product-facing (one question, with your recommended number); from you (stated) if it's internal. **No budget → no stopping criterion → optimization becomes a hobby** — don't start without one. - **Realistic load shape**: the N, the data skew, the concurrency this must hold at (`code-quality.md` → algorithmic sizing). Measuring at toy N is mea