Composer banner restyle performance proof Run on Windows Server 2025, Chromium 156.0.8078.4, 1440x1000, dark theme. Both app versions were installed with vp i and run using vp run dev, on the same leased Windows machine, one server at a time. T3CODE_DEV_SEED=0 prevented copying live data. Each version used its own empty data directory. Set T3CODE_PORT_OFFSET=2200 for base or 2100 for head before starting dev. Set --home-dir to the fresh data directory. From that commit's checkout: node The driver seeds three repositories/projects with four showcase replay threads each through the real app CLI, pairs a fresh browser, then runs three batches of 15 samples at each stage. It reloads a plain thread between batches. The timed function is exactly the PR's 4,000-span probe, returning the numeric milliseconds rather than a formatted string. It never changes loaded CSS rules. Steps per batch: measure plain thread; t3dev.replay('question'); wait for attached banner; measure; select Strict schemas (Recommended); wait until banner disappears; measure; click alpha thread 1 and await its route; measure; reload; measure. Each measurements.json contains every sample and the actual commit and browser version. summarize.py generates timings.png and summary.json from those files. chart.py generates comparison.png. Images of the chat are crops of the same 1440x1000 viewport (x=256..1140) to exclude the machine-identifying environment panel. The composer/banner area (x=380..1116,y=720..982 in original screenshots) is pixel-identical with the question showing. Answered captures show the same submitted answer and provider completion. The two interaction.mp4 files are 20-second excerpts of the same scripted first question/answer/thread-switch sequence. Full runs contained all three measurement batches. Screenshot capture disables animation during the screenshot and requires two identical captures 500 ms apart. It does not alter the CSS during timing. The scenario runner's documentation excludes timing/performance regressions, so this is a manual Windows performance comparison. No unit or browser test result is used as evidence. Not measured: the 14-hover DevTools invalidation trace and exact restyled-element counts; other settings pages and all visual/theme/mobile-width combinations. Verified by gpt-6.1-sol through the Codex harness.