Perf Theory Tester

It enforces a clean methodology: confirm your baseline, apply a single change, run validation passes, then revert before the next test. The output format is tidy, giving you hypothesis ID, change summary, metrics delta, and a verdict with evidence trails. The constraints are strict on purpose, no parallel benchmarks, no stacking changes, which honestly feels rigid but that's the point. If you've ever tried to debug a performance regression where three things changed at once, you'll get why this exists. It's less about automation and more about keeping your experiments from turning into chaos.