[wouterlucas]

What

Adds a cold-startup benchmark alongside the existing throughput (Done!) and memory (Memory!) tests: fresh browser context, load the app, render 1,000 rows, wait for the first renderer-idle, stop the clock. Timing is navigation-relative, so JS download/parse/execute, framework init, font atlas load and scene build are all included (bundle size matters).

Covers all variants: Renderer (+canvasfonts), FTL (+canvas/bitmap/webgl-bitmap), Lightning2 (+core), Blits.

How it works

• Each framework app gets a #startup entry point that renders 1,000 rows once and logs Startup! with the navigation-relative time (Blits gets a /startup route since its router is hash-based)
• New testrunner/runStartup.js: sequential cold loads, fresh context per run, 6x CPU throttle applied before navigation, per-run timeout, all-fail returns startup: -1
• Startup is processed with a plain relative mean vs the fastest framework and reported in its own table section — deliberately excluded from avgMean, so historical Overall scores are unchanged
• writeResults emits a single results_*.html with only the sections that have data, keeping framework columns aligned
• node index.js --startup-only runs just startup + bundle sizes (STARTUP_RUNS, HEADLESS=true supported)

Validation done

• Unit: report pipeline with fixtures (avgMean unchanged, failure renders Failed, no unreplaced placeholders)
• Smoke: Playwright runner against fixture pages (multi-run stats, timeout failure path)
• Real apps: rebuilt Blits and ran the actual startup flow headless (2/2 cold loads reported)

Not included

No generated output (results/, official_results/) and no build side-effects (package-lock link churn) — run on the default benchmark hardware and publish results from there.