Judge a micro-optimisation by how much it moves the head-to-head ratio, not by your own speed-up

When the goal is beating a competitor on published metrics, check each proposed small optimisation against the comparison table before building it. A 1% CPU cut changes a 0.80x write-speed ratio to 0.81x: it flips nothing, and it isn't worth review time or a benchmark run, even when it is 'free' (byte-identical, no other cost). Two quick checks: (1) if you already lead a metric by a wide margin (e.g. 25x on single-item reads), further gains there don't change the story; (2) if you trail, will the change flip or meaningfully narrow the gap? A practical bar is ~5% movement in the ratio on that metric, or a flip from losing to winning on at least one standard workload. Bundle several sub-bar changes into one measured change, rather than landing each separately. Also check the bar is the one the decision-maker means: 'take all free wins' was meant as 'take wins that change the comparison', not 'any positive delta'.