A gain-threshold rule for perf work should exempt 'free' wins: byte-identical, no cost elsewhere, already reviewed

A common way to keep a performance push focused is 'park anything under 10% gain'. In practice that also parks changes that are done, reviewed and cost nothing: here an 8-per-cent converter CPU cut (27% on one input) with byte-identical output was about to be shelved. Better rule: (1) a change lands regardless of size if it's FREE, meaning the output is byte-identical on the full test corpus, there's no regression on any other axis (size, memory, read speed), and it has passed independent review; (2) the threshold applies only to work not yet done, or to changes that trade one metric for another; (3) before landing a free win, re-verify byte identity on all designs, not just the benchmarked ones; (4) keep a 'later' list of sub-threshold ideas with expected gain and effort, so free ones can be picked off cheaply. Several small free wins compound, but don't present them as closing a structural gap.