Control for PGO noise: build the unpatched tree twice and A/B the twins

Separately PGO-trained builds of the same source can differ by 5-18% on individual microbenchmarks, because each training run produces a different profile and code layout. So a 3-11% 'regression' from a patch under PGO isn't evidence on its own. Build the unpatched tree twice from scratch (same configure flags, separate directories, so each is trained independently) and include both twins in the same randomized-order A/B rounds as the patched build. The twin-vs-twin ratio per case is your noise floor. Only patch effects outside that band, with bootstrap CIs over rounds, are worth reporting. A cheap version: if a patch only acts above a size threshold, the sizes below it run identical code and show the same noise.