# Pull request environment attribution: manual proof

Compare base 44a8e9541bc1dc2927944c241b8332125af61757 with head 652dfdcf6193f4495c3934bd536b92e6894faab6. Separate checkouts, fresh databases, Linux Chromium, light appearance, 1280x800 CSS pixels (2x image pixels). Install each checkout with `vp i`.

## Fixtures

Load the authorized T3 Connect test-stage variables into every shell that starts a server or Vite. Use `manual-seed.sh` as the record of the seed commands, with a fresh PROOF_FIXTURES_DIR. Copy forge-fixture.mjs there first. The fixture intercepts only GitHub API calls to return synthetic pull requests. Relay, tunnel, authentication, WebSockets, RPC, server projections and UI are real. No relay or error response is mocked. In the recorded comparison the review server was shared by both clients, and each client had its own primary server and healthy repository.

Start the primary and review servers using the commands at the bottom of manual-seed.sh. Record the web origin printed by the dev runner. Sign in to the stage account in each web app, and dismiss onboarding before recording. The CLI login token cannot be used as a browser token: the real Clerk browser session is required. If automating that login through the loaded SDK, use Clerk.client.signIn.create({identifier,password}), prepareSecondFactor/attemptSecondFactor for email_code when needed, then Clerk.setActive({session:createdSessionId}); credentials come from the caller's stage environment, never this driver.

Evaluate browser-driver.js in each app. Get the review environment ID from its discovery endpoint, then call `await proof.register(reviewEnvironmentId, 'Review server')`. Wait until `proof.state(reviewEnvironmentId).connection.phase === 'connected'` on both sides. The primary environment is labeled Workstation and the review environment Review server. These are synthetic labels supplied through the hostnamectl fixture.

## Recorded interaction, identical on each commit

1. Open /pull-requests?involvement=all&state=open with All servers selected. Both environments have projects, in different synthetic repositories. Confirm the Workstation row and Review server row load.
2. Stop the tracked review server session (SIGINT, and wait for exit). Enter a new search `environment` (the recordings also include a preliminary cached search `healthy`). The relay now returns endpoint_request_failed without a host. Capture 001: base has no environment attribution, head says Review server. Both preserve the healthy row and cached review row.
3. Click Retry while offline. It still fails on both sides; head retains attribution. Retry reconnection is not claimed as fixed. The first pair of recordings contains this interaction; a duplicate still is intentionally omitted.
4. Select Filters > Server > Review server, close the menus, and search `missing-results` to inspect a no-row error. A supplementary run reconnected the same review server on both clients, primed that empty search, then stopped it again. It includes refresh/check-again and another new search `offline-error` before the static captures.
5. For the final no-row still, clear the search and select Filters > State > Closed to exclude cached open rows. Close both menus with Escape twice. Capture 002. This keeps the whole error heading and message visible, with no search activity indicator.
6. Evaluate `await proof.register(reviewEnvironmentId, '')` on both clients. This changes only the actual catalog label; the relay failure remains real. Capture 003: head uses Environment:, base has no prefix. Restore Review server afterward.
7. Select State > Open and Server > Workstation, close the menus. Capture 004: the healthy row remains and the warning clears on both sides.

Each final still was accepted only after two screenshots 500ms apart were byte-identical, then opened for visual inspection. All menus are closed. Recordings are full-tab recordings in two chronological parts per side. The primary recording is unedited; the supplementary recording begins 95 seconds into its raw recording, at the final filter interaction, omitting earlier setup exploration. They show the same actions on each side, with the two tabs driven sequentially.

## Cleanup

Stop all tracked server sessions, then run connect logout with the isolated review base-dir. Remove only the fresh fixture directory and the two proof worktrees.
