# Additional usage-label evidence, round 3

Before: ac5e48c612. After: f77e6dc16f. Product sources were unchanged. All data is synthetic; no application database, live userdata, auth state, or provider was used by this component harness.

Create the additional before checkout with `git -C /home/user/t3code worktree add /home/user/wt/pr-32534-before3 ac5e48c612`, then `vp install` there. The after checkout is `/home/user/wt/pr-32534`. Keep all drivers in this scratch directory and symlink its `node_modules` to `/home/user/wt/pr-32534/apps/web/node_modules`.

Reuse recapture2's harness and capture/diff/sheet drivers. `driver-changes.diff` lists every change: scratch paths, private capture ports 61334/61335, A5/A6/T1/D1 synthetic fixture branches, a T1-only wrapper width, the requested capture matrix/provider heading, and sheet row selection. The diff generator and index are byte-identical to recapture2. Resize drivers derive from capture's original drivers and replace CDP transitional screencast frames with settled screenshots held for 2.5 seconds each.

Start in retained terminal sessions, recording the session ids:

```sh
T3CODE_DEV_SEED=0 node /tmp/lead-32534-r1/capture3/harness-server.mjs /home/user/wt/pr-32534-before3 61335
T3CODE_DEV_SEED=0 node /tmp/lead-32534-r1/capture3/harness-server.mjs /home/user/wt/pr-32534 61334
```

These are Vite component-harness servers, not T3 backend servers. Each imports its own checkout's real UsageLimitsPooled, WorkspacePageContainer, CSS and Inter. Neither creates or seeds a database. Do not set VITE_HTTP_URL or VITE_WS_URL.

Run `node capture-harness.mjs` then `node record-resize.mjs`. The static driver creates a timestamped run and latest-run.txt. The resize driver writes only to that run. Run `make-diffs.py <run>`, `encode-resize.py <run>`, and `prepare-sheets.py` with the preserved capture virtualenv's Python (Pillow 13.0.0, NumPy 2.5.4). Run `node probe-layout.mjs` for independent geometry measurements.

Conditions: Playwright 1.60.0, Chromium 148.0.7778.96, DPR 1, Inter Variable, 16 px root, en-US/UTC, light theme, reduced motion, time frozen to 2026-10-25T16:00:00Z. Before each capture: network idle, both Inter font weights ready, cursor outside the card, and two byte-identical screenshots 500 ms apart. All static and resize frames settled on the first attempt.

A5: Work ChatGPT, no reset and no credits. A6: long display name, no reset, 12 credits. T1: same long display name, reset 2026-10-26T15:59:00Z (23h 59m), 12 credits; T1-only wrapper is 954 px or 953 px inside the unchanged 1280×900 viewport, yielding measured pool widths 672 px or 671 px respectively. D1: driver claudeAgent, with email-only and no-name/no-email variants. All windows are weekly and 40% used. A1 is the original email-only Codex fixture, unchanged.

Resize schedule in each encoded 7.5-second video: 0.000–2.500 s = 1280×900/pool 742 px; 2.500–5.000 s = 390×900/pool 316 px; 5.000–7.500 s = 1280×900/pool 742 px. Each actual browser step settles first, then holds unchanged for at least 2.5 seconds, with another byte-equality check at the end. The actual wall-clock resize/settle/hold timestamps are in the respective schedule JSON files. The videos repeat only those three settled PNGs, each for 2.5 seconds; waiting and transitional frames are excluded. Compact screenshots retain their real 390 px width and are padded on the right to the fixed 1280 px video canvas, rather than stretched.

Diff0 marks every pixel with any channel delta >0 as full magenta; DiffT does the same for >16. The original full viewports, per-row full-viewport contact sheets, readable detail crops and per-side resize storyboards are retained. All were visually inspected before upload. Compact A1 and T1 at 671 px are byte-identical before/after. Every nonzero difference is within the added account label. The A6 credit pill is 40 px on both sides; the T1 wide reset/credit pill is 115.21875 px on both sides, with its credit child inside its right boundary.

Upload only the allowlisted synthetic evidence and drivers with `ghdrop upload -m`. No runtime model identifier is included. The evidence markdown is ready to paste; the PR body was not edited. Stop only the retained sessions and remove only `/home/user/wt/pr-32534-before3` when done.

AI-assisted; verified with the tests and tools listed above.