Windows branch-picker evidence

Surface: production web bundle served by the real Node server on Windows Server 2025.
Git 2.56.1.windows.1; Node 26.13.1; Chromium through Playwright 1.64.1.
Light theme, viewport 1600x1000, same leased machine and synthetic repository.
Base: 1458a4241939b9ba17f78dc219205d1a3fc64557
Head: 108b60ab85b366e97a5bca0c2c36a4978254f7a3

Create worktrees at C:/p/base and C:/p/head at these commits.
Run vp i in each, then vp build in each apps/web directory.
setup.ps1 records the setup commands used after base dependencies had already been installed.
It builds both sides and creates a bare local remote and a 33,232-file clone on work,
plus feat-a, feat-b and feat-c branches pushed to the remote.
Set synthetic git user.name and user.email on C:/p/repo.

For each side, sequentially (never run both servers together):
  powershell -File start.ps1 base
  Wait until server is listening on 127.0.0.1:4800 and provider discovery is complete.
  powershell -File seed.ps1 base
  Wait for three replay turns to complete.
  $env:PW = '<installed playwright package directory>'
  node measure.cjs base
  Stop the process tree at the PID written by start.ps1.
Repeat with head. Each gets its own fresh T3 home, with dev seeding disabled.
Do not copy a live database. No provider credentials are needed.

measure.cjs records all nine unchanged opens, then creates external-proof with git,
opens the picker again, records the new option and removes the external branch.
First open is retained but excluded from steady-state metrics because it establishes
head's refs digest. The following eight runs are the steady-state comparison.
The script writes each run into its own timestamped results directory.
It records the complete UI video and uses two identical screenshots 500 ms apart.

Copy that side's results and server trace files locally, then run:
  python3 analyze.py base <results>/raw.json <trace-directory> base-measurements.json
  python3 analyze.py head <results>/raw.json <trace-directory> head-measurements.json
  python3 report.py
report.py uses Pillow to render the measurement output. The two JSON reports contain
per-open sanitized RPC sequences and git spans; paths and trace attributes are omitted.
The list reload metric is the duration of the listRefs RPC root which owns its git work.
Background subscription and diff-preview git work is reported but excluded from picker counts.

Recordings were trimmed by 20 seconds to omit pairing and initial navigation, encoded
as MP4, and the machine label was masked at x=1346,y=75,w=110,h=31. App stills have
the same label masked. No branch controls or measurement output were altered.
The uploaded video retains the scripted open/close cadence and external-branch check.

Results: every steady base open sent two refresh:true listRefs requests and ran six
list reload git processes, median 180.545 ms, range 150.22–223.95 ms, spread 73.73 ms.
Head sent none, with zero reload work. Picker-related total git fell from 16 to 12:
head's status refresh includes two additional ref-snapshot reads. Both sides displayed
the same branch options, and external-proof appeared on the next open.
Visible cached-list latency did not improve (base median 64 ms, head 77.5 ms).
The result proves removal of redundant server reload work, not a visible-list speedup.
