# Divider-prompt renderer performance proof

Two independent checkouts run the real web client and server. The base is the merge-base of the proposed commit with main. Both sides use Linux x64, Node 26.13.1, Chromium headless, a 1440x900 viewport, and dark theme. Run only one measurement process at a time.

1. Install each checkout with `vp i`. Start its server once with `T3CODE_DEV_SEED=0 vp run dev` to create its fresh database, then stop it.
2. Add a synthetic project through `node apps/server/src/bin.ts project add <synthetic-git-repository> --title 'Divider performance demo' --base-dir .t3` in each checkout.
3. With that server stopped, run `node seed.mjs <checkout> <checkout>/.t3`. This uses the checkout's schema decoder and server EventSink.write service to commit synthetic historical provider-native Codex children, messages, timeline items and completed root nodes. It neither edits SQL rows directly nor launches a real agent. The opening messages deliberately omit workflowScript, matching historical threads before that flag existed.
4. Restart each checkout with `T3CODE_DEV_SEED=0 vp run dev`. Read its actual web port. Run `node measure.mjs <checkout> <web-port> <base-or-head> <output-root> 7` once per side. Pairing credentials stay in memory and recording begins after pairing.
5. The driver intercepts only the served subagentRuntime module to add performance.now immediately around the unchanged isWorkflowScriptMessage body. It checks that real ChatView rendering called it, every result is false, every flag is absent, and the ordinary prompt is visible. It does one warm-up plus seven full navigations for each identical input. Profiling covers the first measured navigation at a 1ms sampling interval. Every screenshot must be identical to another 500ms later. Both screenshots and videos crop the environment sidebar containing the machine label; the browser viewport itself is unchanged. CPU profile source paths are anonymized.
6. Run `node node-measure.mjs <checkout> <base-or-head> <output-root>` sequentially for supplemental exact-expression Node timings. It extracts the expression from that checkout, records one warm-up and seven measured rejections, and tests the two large stress inputs on head only.
7. Run `node report.mjs <checkout> <base-app-output> <head-app-output> <base-node-output> <head-node-output> <report-output>` to combine raw measurements and capture the readable output. comparison.json contains all raw samples, cold/warm-up results, classifications, metadata, medians and spreads.

Visible-prompt time ends when the browser driver acknowledges the actual prompt; it is not a first-paint timestamp and includes inability to acknowledge a visible prompt while the renderer is blocked. Detection totals include calls through the following renderer acknowledgement, so late render work may extend beyond first visibility. Run counts differ because slow rendering changes the number of React updates; each actual call is retained. The per-opening totals and full navigation times measure the resulting cost in the real app rather than forcing an artificial fixed call count.

The published videos are separate single-opening recordings from record-opening.mjs, on the unmodified app. They show the same parent-to-child navigation on both sides, and wait until syncing is gone and two screenshots 500ms apart are identical, then retain another 1.5 seconds. The only edit is cropping out the environment sidebar, without changing playback speed. The seven-run measurements and CPU profiles come from measure.mjs. Profiles can be loaded in Chrome DevTools' JavaScript Profiler. The final proof uses seven-run medians and conservatively compares the improvement to the sum of both sides' full max-minus-min spreads. The three-divider navigation is also reported even if normal app startup noise hides its smaller detector improvement.

No tests are used as performance evidence. No source changes or evidence files are committed. The scenario runner was not used because its documentation explicitly excludes performance regressions from its proof mechanism.
