Branch-list reload performance proof

Platform: Windows Server 2025, 4 vCPU, 16 GiB RAM, Git 2.56.1.windows.1,
Microsoft Defender antivirus and real-time protection enabled.
Chromium headless, Playwright 1.64.1, 1600x1000, dark theme.

Base: 997575be5f5dde8e5443a26ae4cfd4155fa27d87
Head: 09d4ab369fc338b7b44098a0a0d51d5073260c6b

Build each revision in its own worktree under C:/p/base and C:/p/head.
Run vp i in each root, then vp build in each apps/web, sequentially.
Set T3_SOURCE_CHECKOUT to a synthetic source checkout before seed.ps1.
seed.ps1 creates the bare remote and synthetic repository plus separate app homes.
The checked-out benchmark repository has 33,231 tracked files and five local branches.
Only one app server runs at a time, port 4800. No install/build runs during timing.

For each side, run run-server.ps1 <base|head>, wait for port 4800 to listen,
then seed-app.ps1 <base|head>. It uses the free command-output replay agent.
Set PW to the absolute directory of the installed playwright-core package.
Run node C:/p/measure.cjs <base|head>.
Then node C:/p/analyze.cjs <base|head> and node C:/p/export-trace.cjs <base|head>.
Stop the server process tree using the PID captured by run-server.ps1 before the next side.
The driver writes a new timestamped output directory on every invocation.

The measured action is opening the Branch combobox, waiting 3 seconds, pressing Escape,
and waiting 3 seconds. There are 21 opens, analyzed as three consecutive runs of 7.
Reload wall time is the interval from the earliest listRefs request start to the latest
listRefs request end for that open, from real server trace spans, as in the reporter's driver.
This is server RPC wall time, not click-to-paint latency.

All samples are retained, including outliers. Run spread means max minus min of the
three 7-open medians; individual-open ranges are also reported separately.
Each side additionally runs git checkout feat-a outside the app and reopens the picker.
The settled screenshot and final driver observation verify the current branch changes.
Screenshots are saved only after two captures 500 ms apart are identical.

The initial base run is retained as base-measurement.json. It read the post-checkout
trigger label before the final settled capture, so its metadata says work even though
its terminal-checkout screenshot shows feat-a. The final base repeat and head use the
identical final driver, which reads that label after the settled capture.

Sanitized traces preserve only span names, IDs, parent IDs, timestamps and durations.
They contain no attributes, commands, paths, credentials, or machine identities.
No unit-test results are used as evidence.
