extremely-optimize

# Extremely-Optimize Patching a slow subsystem keeps the shape that made it slow. Each fix threads through the structure responsible for the cost, so the structure survives the fix and the next fix buys less. This skill stops patching: measure what the work must cost, tear out the implementation that misses it, and rebuild against that number. The number is the **floor** — the work the contract cannot avoid. Bytes that must move at achievable bandwidth. Comparisons the lower bound demands. Round-trips the protocol requires. Every target carries a multiple of its floor, and that multiple, not intuition, decides what gets rebuilt and when to stop. Order is fixed. **Hot** first, because a factor there is the whole win. **Cold** after, because cold code is never free — but it is graded in a different currency, and speed is rarely it. ## Scope Repo-wide is opt-in, never inferred. - A named path, or a target identified in the request, is the whole job. - A bare invocation, or wording that asks for a sweep, profiles the repo's own workload and works the ranked list. A campaign runs one target at a time, each landing as its own atomic commit. Half-rebuilt is the forbidden state: finish a target or revert it. ## Workflow 1. **Pin the workload and the baseline.** Find or write one runnable command that exercises the target with representative input and takes at least a second of wall clock. Run `hyperfine --warmup 3 --min-runs 10 '<cmd>'`. Record median, stddev, min, max. ...