Whole-project, profile-driven performance campaign that establishes durable benchmarks and folds proven optimizations back in place — build macro benchmarks for the flows that matter (committed to the repo as a first-class deliverable), profile them to rank where the time actually goes, attack only the dominant vectors (the smallest prefix of the ranking that dominates the profile, cap 3 per iteration, each carrying its measured share and Amdahl ceiling — a ceiling inside macro noise is not a vector, and speculation never is), pin each vector's contract with a differential + property harness under the PINNED/FLEXIBLE/SUSPECT split before anything races (SUSPECT behavior a…
