Manual real-app proof

Environment: Linux x86_64, web client in Chromium, 1280x800 CSS viewport (2x capture density), light theme. Separate fresh homes and Git repos per commit. Replay (dev), multi-turn recording, one completed turn per thread. No live database copied.

1. Start each checkout with T3CODE_DEV_SEED=0 and SHELL=/bin/sh. The current-main scenario start-app helper was used to prepare synthetic settings, disable paid providers, and run the actual checkout's server and Vite client. Its environment label shim displays Scenario.
2. Run `node seed.mjs CHECKOUT FRESH_HOME STATE_JSON`. It adds a synthetic Git project through the CLI, completes Replay turns for A and B, and seeds one legacy ref per thread plus an unknown-owner ref. Each thread has two actual v2 refs and one synthetic legacy ref.
3. Run `node report.mjs STATE_JSON`. It writes ref-report.py into the synthetic project; that script reads current refs directly with git for-each-ref, using the actual seeded ownership lists.
4. Pair the browser and open A by its environment/thread URL. Open its terminal and run `clear; python3 ref-report.py` to see all seven refs.
5. Record the browser. Right-click A's sidebar card, Delete, Confirm. Wait until its card disappears and its projection tombstone is persisted.
6. Run `node restore-backlog.mjs STATE_JSON` after deletion, preserving refs for a deleted owner just as an older install would. Restart the real server against the same home: the development Node watcher was triggered by touching storageCleanup.ts without changing its contents. This exercises the real startup sweep.
7. Open Settings > Maintenance. Enable Delete old rotated logs; press Enter to save the initial 8-day draft. Click Delete now and wait for Last run just now.
8. Open surviving B by URL, open its real integrated terminal, and run `clear; python3 ref-report.py; printf '\033[?25l'`. Capture the current Git ref counts. Stop recording.

Observed on base 9ab1d22d8fc0bbf491d92995f92fb6eb94efe8b0: deleted A retains 3/3 refs (2 v2, 1 legacy); active B retains 3/3; unknown owner retains 1/1.
Observed on head 26842b782e17e081d5829835378fd45441706fd7: deleted A retains 0/3 refs after restart and after manual cleanup; active B retains 3/3; unknown owner retains 1/1.

The maintenance report counts filesystem cleanup items, not checkpoint refs; both sides can say 0 removed, 1 kept. The terminal reads the actual Git refs and is the verdict evidence.

Each final still was accepted only after two screenshots at least 500ms apart had zero changed pixels. All four stills and both video contact sheets were visually inspected before upload. The videos contain the real deletion, restart/reconnection, manual maintenance action, and terminal readout; their different idle durations reflect manual driving, not a performance comparison.

Scenario attempt: checkpoint-ref-cleanup.e2e.ts was tried first. After three corrections it still stopped at a terminal-canvas pointer interception, before its ref assertion. It is supplied for transparency, not as a successful scenarios:prove verdict. The manual run above supplies the proof.
