Web performance audit

Field measurement from a real browser, not lab guesses. The goal is usually LCP — either lowering it, or discovering the reported number is not measuring the thing anyone looks at.

Attribute before you infer. Every wrong answer recorded at the bottom of this file came from reasoning about a plausible cause instead of asking the browser which element, which frame, which importer. Attribution is nearly always one observer or one flag away.

The loop

1. Capture a trace (below), or take one you have been handed.
2. Read it with the analyser: pnpm perf-trace summary <trace.json>.
3. Attribute whatever looks wrong — element, frame, importer — before proposing a fix.
4. Change one thing.
5. Re-measure and compare: pnpm perf-trace compare <before> <after>.

Stop when the number moves or the evidence says the number was never the problem.