Measure, don't guess — profile before you optimize

Any performance work, on any platform. The instant the question is "why is this slow?" the answer is "profile it," not "I bet it's X."

Human intuition about performance is notoriously unreliable: the code that looks expensive often isn't, and the real cost hides somewhere unglamorous (a redundant load, a cache-hostile loop, an accidental per-frame allocation). Optimizing without measuring means you spend effort on code that wasn't the problem, complicate the codebase, and frequently don't move the frame rate at all [S-perf-profiling]. A profiler replaces the guess with a fact — this function, this system, this draw call is eating the frame — so every hour of optimization is aimed at something that matters. "Measure, don't guess" is the same discipline as judging design by observed play (DESIGN-0001) and player behavior (PLAYTEST-0001), applied to performance.

Profile on representative hardware and content, not a beefy dev box on an empty scene. Identify the single biggest cost, fix it, then re-profile (the bottleneck moves — see PERF-0003). Distinguish CPU-bound from GPU-bound before choosing a fix (PERF-0004). Keep before/after measurements so you can prove a change helped rather than assuming it did.

No serious dissent: profile-first is near-universal engineering practice. The only nuance is how much "obvious" optimization to do without formal profiling — and even there, the safe answer is to verify.

The foundational PERF principle and the resolution of the PERF-0001 forward-reference from PLAYTEST-0005 (telemetry/measurement). Shares the "measure, don't guess" ethos that runs through DESIGN-0001, PLAYTEST-0001/0005, and SYS-0002. Confidence 5.