Performance: measure, find the real bottleneck, fix the biggest, verify

Slowness has a real, measurable cause. The cardinal sin is optimizing by intuition. You'll spend effort on the wrong thing and maybe trade away correctness for nothing. Get data first.

The loop

1. Reproduce + measure. Get a real number under a realistic scenario: wall-clock, FPS/frame time, query ms (EXPLAIN ANALYZE), request latency, bundle size, memory. No number, no optimizing. Write it down; it's your before.
2. Find the bottleneck: profile it. Use the right instrument (browser Performance panel / React Profiler, a flame graph, DB query plan, a tracer) and find where the time actually goes, the ~20% causing ~80%. The universal classes of waste: repeated work (recomputed per item/render instead of once), amplified work (one request fanning out into N), missing lookup structure (a scan where an index/map belongs), serial waiting (round-trips that could be batched or parallel), oversized payloads, and no caching of stable results. For the domain-specific suspects, load the active field pack (engineering/active-field.md → the pack's performance section); if the field has no pack, let the profile, not a checklist, name the suspect.
3. Fix the biggest one. Make the single change with the most impact; resist micro-optimizing noise. Prefer doing less work (cache, batch, index, memoize, defer, paginate) over doing the same work faster.
4. Verify the win. Re-measure the same way: confirm the number actually moved, and that behavior and correctness are unchanged (core/rigor.md). A "faster" version that's subtly wrong is a regression.
5. Guard it. Note the metric (a comment, a budget, a perf test) so the regression is visible next time.

After

Run levelup (capture) to record the bottleneck class and its lesson in the active field's lessons.md: including the wrong suspect you ruled out, so MasterMind doesn't re-profile it next time. Report: before → after numbers, the cause, the fix, and the guard added.

Gotchas

• Measure before and after: "feels faster" is not a result; a number that moved is.
• Profile over pattern-matching. The obvious suspect is often not the bottleneck, the profiler decides.
• Correctness is not on the table. Never trade a correct result for speed; if a fix changes behavior, it's not a perf fix.
• Biggest first. One 10× hotspot beats ten 5% tweaks; stop when the number is good enough (match effort to stakes), not when you've micro-optimized everything.
• Not debug (a wrong result) and not qa (proving it works). This is make the correct thing fast.