measured-performance-optimization — measure, change one thing, prove it, or roll back

YOU MUST EXECUTE THIS WORKFLOW. Do not just describe it. This skill is a disciplined optimization loop, not advice. Every iteration produces evidence.

⚠️ Critical Constraints

• No measurement, no change. You may NOT edit code for performance before a baseline profile + benchmark artifact exists on disk. Why: un-profiled optimization is guessing; you cannot prove improvement against nothing, and you will "optimize" code that is not the bottleneck.

  • WRONG: // looks slow, let me rewrite this loop then commit.
  • CORRECT: Run scripts/bench.sh baseline → read profile → target the top frame.

• One change per iteration. Apply exactly one optimization, then re-measure. Why: batched changes make a regression impossible to attribute, so you cannot roll back the one that hurt while keeping the ones that helped.