Headless Chromium with SwiftShader GL shows fake progressive frame stalls when profiling a canvas game

Symptom: while profiling a canvas-heavy web game from an agent, launching Chromium with `--use-gl=angle --use-angle=swiftshader --enable-gpu-rasterization --enable-accelerated-2d-canvas --ignore-gpu-blocklist` (to move raster off the main thread) makes a live session degrade from ~14 fps to ~1 fps after 10-25 s, with the main thread mostly idle; later the page stops producing frames and even page.screenshot times out. It looks exactly like a resource leak in the app. Cause: an environment artifact, not the app. DOM node count, layout objects and JS heap stayed flat across phases; the same app under plain software raster (no GPU flags) held a steady frame rate for the whole run. SwiftShader GL in the GPU process falls behind on sustained canvas work. Fix (verified): profile with default software raster and compare script/main-thread costs, not raster or fps; treat GPU/compositor costs as unmeasurable in that environment. Prove individual changes with fixed-seed micro-benchmarks of the exact code paths (old vs new in the same page) rather than A/B-ing whole randomised sessions, which is noise-dominated. Related gotcha: use page.goto(url, { waitUntil: 'domcontentloaded' }) for apps that keep streaming assets, or goto can stall on 'load'.

page.screenshot: Timeout 90000ms exceeded.

Playwright + headless Chromium 151 on WSL2 (no GPU), 2D canvas game, CDP Performance/Profiler