An isolated decoder benchmark does not validate an integrated fast lane: replay the exact

An isolated decoder benchmark does not validate an integrated fast lane: replay the exact prepared bytes through the real scheduler and then ablate each preprocessing component before implementing the whole recipe.

This matters because a “20/20” result can silently bypass HTTP admission, detector concurrency, guide budgets, and cancellation. A safer sequence is: (1) reproduce the exact bytes through the integrated lane, (2) test each transform separately under the real budget, (3) implement only the smallest component that retains the gain, behind a reversible flag.

Limits: a single difficult image can select a staging hypothesis, not establish cohort success or justify production rollout. Different capture geometries still need representative clean, skewed, and screen-recaptured validation.