Performance proof driver (Windows Server 2025, Chromium 156.0.8078.4, Azure CLI 2.92.0 with azure-devops extension 1.0.10) 1. Export AZURE_DEVOPS_EXT_PAT, AZURE_DEVOPS_TEST_ORG and AZURE_DEVOPS_TEST_PROJECT for the dedicated test organisation. 2. Run seed.py on the VM. It creates ten PRs with the same four changed files and writes fixture.json. Keep fixture.json private: it identifies the test organisation. 3. On one Windows lease, install base d080e4eb587898c5057d03b3c7148d113ddeccc3 and head 3900e23b57214590dd654ff33b6825aa677ae5a6 in separate git worktrees, using vp i in each. Install Azure CLI and its azure-devops extension. 4. Copy fixture.json to the Windows scratch directory. Create private-env.json with the three Azure variables and a random T3CODE_DEV_AUTH_TOKEN (64 hex characters). Never upload that file. Initialize scratch/repo with the fixture's remote and fetch/check out its synthetic base branch, using an environment-backed credential helper. Configure az's organisation and project defaults. 5. Add that repo as a project on each side's isolated home and record its returned id as [{"id":"..."}] in project-base.json / project-head.json. start-windows.ps1 illustrates the commands; the capture used the equivalent commands with absolute paths in its private launcher. Start each dev server with T3CODE_DEV_SEED=0, no automatic cwd project, the reusable token, and port offset 2000 for base / 3000 for head. Retain the process ids for cleanup. 6. Run node measure.cjs base , then node measure.cjs head . Do not run installs or other workloads during measurement. Each side gets a new browser context, dark theme, and 1440x1000 viewport. Each run navigates to one of the ten PRs, waits for Summary plus six seconds, clicks Code, waits for g.ts, verifies all four paths, captures the stable diff, and opens Timeline. The first three runs are warmups; all remaining seven are measured. There are no injected provider delays or mocked responses. 7. Copy the timestamped result directories back to the VM. Run render-evidence.py with ffmpeg and matplotlib available. It renders unchanged-speed clips of every measured interaction, matched run-10 stable screenshots, all numerical results and a chart. The films crop away the sidebar and apply account masks throughout each entire clip. The browser driver additionally redacts email text before paint. Cuts between runs remove navigation/loading of the next PR; each measured Code interaction remains continuous. 8. Stop the servers by their retained process ids. Run seed.py --cleanup with the same Azure environment to abandon just these ten PRs and delete their eleven branches. Release the Windows lease and remove the isolated worktrees and scratch material after uploading the safe evidence. The separate redact-recording.py repairs the prior rejected recording; it does not alter timestamps or measured values. Its author mask is enabled from the first frame through the last. That repaired prior recording is supplementary to this fresh comparison.