01
Reproduce
Make the expensive interaction short, deterministic, and easy to repeat before collecting a trace.
02
Locate
Separate main-thread work, allocation pressure, drawing cost, and GPU composition instead of treating them as one bottleneck.
03
Confirm
Compare the same frame window before and after the change; smoothness claims need comparable evidence.
Working rule
Make the next measurement cheaper than the next argument.
Keep a minimal reproduction close to the observation. Change one variable at a time, preserve the failing path, and prefer before-and-after evidence over a broad rewrite.