GitLab watch performance proof — real app measurement

Environment: Linux x86_64, Node v26.12.1, glab 1.124.0; web clients Chromium,
1280x800 CSS pixels, light theme. Both servers run simultaneously on the same
machine, from separate worktrees and fresh databases. Their sweeps are staggered
by about 21 seconds. No test doubles, accelerated timers, or production state.

Base: 9db3cb10916c81830efa0d0d7be6601d2fe51158 (merge-base with origin/main).
Head: 8ec03413488d284ca88af77a5ef01c9a854fd9f8.

Reproduction:
1. Install each checkout with `vp i`. Start `T3CODE_DEV_SEED=0 vp run dev` in each.
2. Run fixture.py once in a scratch directory. It creates a private synthetic
   gitlab.com project, feature branch, MR !1 and running external ext-ci status,
   then clones the fixture repository. Keep fixture-private.json private.
3. In each checkout, use the app CLI to add the clone and start a free Replay thread:
   node apps/server/src/bin.ts project add <clone> --title 'Watch proof' --base-dir .t3
   node apps/server/src/bin.ts thread start <clone> 'Watch proof baseline' \
     --title 'GitLab watch proof' --provider devReplay --model multi-turn \
     --wait --json --base-dir .t3
4. Pair each dev browser, open the CLI-returned thread by its environment/thread
   URL, and run start-watch.js through preview_evaluate. This invokes the same
   client command as Watch for changes, linking MR !1 and enabling its watch.
   Both show Waiting · Watching pull request #1. Close foreground MR detail views.
5. Set timeline.py's `start` to the shared watch start epoch, then launch:
   python3 timeline.py <head-checkout> <base-checkout>
   It keeps the check running for 20 minutes, changes it to success at +20 min,
   posts a viewer comment at +23 min, updates the remote feature branch at +37 min, and stops at +47 min.
6. Run `python3 report.py <run-directory>` to render the measured 20-minute output.
   measure.py walks trace span ancestry; each VcsProcess.run under a watch sweep
   is one subprocess. The GitLab watcher performs glab reads, including an
   un-named viewer lookup. Link-sync and foreground UI requests are separate.
   The initial baseline sweep is included in totals but excluded from medians.
7. Before publishing, inspect every image and representative video frames, and
   upload only synthetic measurement files. Do not upload fixture-private.json,
   raw traces, dev logs, or uncropped recordings. UI context videos crop away the
   sidebar, environment label, and filesystem controls.

Measurement scope: repeated normal two-minute sweeps, not independent two-hour
watch sessions. API caching can reduce some base reads; all observed calls are
retained. Sweep duration is wall time of PullRequestWatchReactor.sweep, not summed
parallel subprocess duration. No projected two-hour number is presented.

Clean up: stop both dev process trees using the retained terminal sessions;
DELETE the synthetic GitLab project via glab api projects/<id> -X DELETE; remove
only the task-created worktrees and scratch directory after uploading evidence.

After the scripted +47-minute sequence, observe.py extends collection through
+55 minutes to capture the periodic full refresh, without modifying GitLab.
Command: python3 observe.py <run-directory> <head-checkout> <base-checkout>
