# Manual reproduction

Run each commit in its own checkout, install with `vp i`, and use a fresh `.t3/manual` home. Never seed from an existing T3 database.

Start the synthetic API with `GITHUB_OUTAGE_FIXTURE_DIR=<scratch-directory> python3 fixture.py`. In a separate `GH_CONFIG_DIR`, write `config.yml` with `http_unix_socket: <scratch-directory>/github.sock`. Run both apps with this configuration and a synthetic GH_TOKEN, plus `T3CODE_DEV_SEED=0` and `T3CODE_AUTO_BOOTSTRAP_PROJECT_FROM_CWD=0`:

```
vp run dev --home-dir .t3/manual --no-seed
```

Create the synthetic repository using `python3 seed.py <workspace>`. In each checkout register it with:

```
node apps/server/src/bin.ts project add <workspace> --title search-outage --base-dir .t3/manual
```

Pair a browser with each server. Use a 1280 × 800 CSS viewport, light theme. Open `/pull-requests` with no query and press the button with accessible name `Refresh pull requests`. Wait for that button to be enabled again. Base shows no PRs; head lists the synthetic open PR. Two captures at least 500 ms apart were pixel-identical on each side before keeping the still.

The recorded interaction on each side is identical:

1. Click `Refresh pull requests`, then wait until enabled.
2. Fill the `Search pull requests` searchbox with `outage`, press Enter, and click refresh again.
3. Wait until refresh is enabled, then capture the query state.
4. Clear the searchbox, press Enter, refresh and wait until enabled.

Each recording is approximately six seconds. Query captures also passed the 500 ms pixel-stability check. The head keeps previously loaded matching rows while the query fails. The actual WebSocket responses still contain `scenario/search-outage could not be read.` with an empty entries array; this is preserved in the synthetic-data transcript. No network or application response was replaced in the browser; only the upstream GitHub endpoint was controlled.

The attempted scenario was abandoned after three locator/state corrections. Its assumption that a failed text query always renders `Nothing matches` is false on the head because the client retains matching baseline rows. The manual captures, not a scenario pass/fail verdict, are the proof.
