Performance loop

1. Choose a stable benchmark and preserve its identity and workload.
2. Record a statistically useful baseline.
3. Profile the hot path before editing.
4. Implement one bounded optimization.
5. Re-run the same benchmark and correctness checks.
6. Keep the change only when the improvement is reproducible and no contract regresses.
7. Use a new benchmark identifier for a materially different workload.
8. Run review.

Do not report performance gains from debug builds or incomparable fixtures.