## Environment and method

Both recordings show T3 Code's real web client in Chromium on Linux, dark theme, a 1280 × 800 CSS viewport. Each commit ran with its own installed dependencies, dev server, worktree, and fresh database with automatic seeding disabled. No live user database was copied. No app source was changed.

Before: [d290c1f1b6](https://git.cafe/pingdotgg/t3code/commit/d290c1f1b6c93416dfda6c9f7cd15d928a9aeb4c).
After: [515dc6eba6](https://git.cafe/pingdotgg/t3code/commit/515dc6eba667cc9346e17f9060ba72e089cb23c2).

The seed creates identical synthetic Git repositories and two empty threads through the server's public HTTP/RPC APIs. The recorded interactions use the real branch selector and sidebar thread buttons. No agent turns run.

External forge data is synthetic: forge-preload.mjs redirects only requests for the synthetic evidence/branch-demo repository to forge-fixture.mjs. All other fetches retain normal behavior. The fixture holds the forge list response until after navigation has committed a visit. This controls network timing without changing the app's discovery, orchestration, persistence, or rendering code.

## Same steps on both sides

1. Open Branch PR discovery on feature/alpha and confirm synthetic PR 101, Feature Alpha.
2. Arm the fixture with GET /arm. Use evidence.switchBranch('feature/beta'), which clicks the app's branch selector and its option.
3. Run evidence.visitDuringLookup(): click Notes, wait 700 ms, click Branch PR discovery, wait 700 ms.
4. Confirm the fixture has a pending response through GET /status, then release it with GET /release. Wait 3 seconds and read evidence.state().
5. Repeat steps 2–4 with feature/no-pr.

Before: feature/beta retained PR 101; feature/no-pr also retained PR 101. The server logged rejected thread.pull-request.sync commands following both intervening visits.

After: feature/beta changed to PR 202, Feature Beta; feature/no-pr cleared the PR.

The videos preserve the full interactions. They were resized from the recorder's 2× pixel output to the CSS viewport resolution. Only the machine label was covered; branch names and PR state were untouched. Both recordings were inspected across their timelines and at the result frames before uploading.

## Driver setup

Run forge-fixture.mjs with Node. For each checkout start the ordinary dev server with T3CODE_DEV_SEED=0, NODE_OPTIONS pointing to forge-preload.mjs with --import, and an arbitrary synthetic CAFE_TOKEN value. Pair the browser using that server's startup pairing link. Run seed-repository.sh with a new, timestamped output directory, evaluate browser-driver.js in the authenticated browser, call installEvidenceDriver(), then evidence.seed(workspace), where workspace is the new synthetic repository directory. Use a fresh database and new repository directory for a repeat run to avoid cached lookup results.

No tests are offered as evidence. Desktop shell behavior, mobile rendering, uncontrolled live-forge latency, and linked-PR replacement guards were not exercised.

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