Branch-picker Windows performance evidence

Base: 755de001f25c54c9d6435bb23a83ffcd231e06a5
Head: b88235da13b638c86634299a8d2f17c67b824dca

Manual harness: docs/operations/scenarios.md explicitly excludes timing proofs.
One leased Windows Server 2025 Datacenter machine, version 10.0.26100,
4 logical CPUs, 16 GB RAM, Defender real-time protection enabled,
Git 2.56.1.windows.1, Node v26.13.1, headless Chromium.
Built web clients served by their own real Node server, one at a time.
Fresh database per revision, T3CODE_DEV_SEED=0, synthetic replay thread.
The same 33,231-tracked-file repository and local bare remote on both sides.
1600x1000 viewport, dark theme, thread details panel visible.

Driver: common.cjs and branch.cjs, adapted from the issue's verbatim driver.
Added WebSocket Exit timestamps, video after pairing, outside-checkout check.
RUNS=7 SEQ=1; alternating feat-a/work starting on work; branches feat-a/b/c
are HEAD~20, all four branches have upstreams. Picker waits 2.5 seconds before
clicking the option, then observes six seconds after the label updates.
Finally git checkout feat-b outside T3; wait for picker to report feat-b.

Primary performance metric: time from clicking the option until the last
completion of any vcs.* RPC sent after that click, within the observation
window. All measured VCS RPCs completed (pending arrays empty).
Base completion times: 5476,4874,4192,5162,5371,5324,5254 ms.
Head completion times: 2939,2844,2998,3050,3416,2825,3036 ms.
Medians 5254 -> 2998 ms; reduction 2256 ms (42.9%).
Base spread 1284 ms, head spread 591 ms. Reduction exceeds their sum 1875 ms.
Ranges do not overlap. Seven successful branch changes on each revision.

Base: extra client refreshStatus localOnly requests after 5/7 switches.
All five corresponding server root spans ran 10 git processes.
Extra server durations: 2121.433,2728.8361,3003.9073,2773.1685,2831.4175 ms.
Head: zero extra refreshStatus requests after 7/7 switches.
Refs requests per switch fell from 5 to 3.
Whole observation git medians 98 -> 94; ranges 77-125 -> 92-118.
Do not interpret noisy total git counts as a precise ten-process saving.

Outside checkout: base picker updated in 1976 ms, head in 1916 ms.
Both sent one refreshStatus for this external checkout.
Visible picker-label latency medians 2388 -> 2373 ms; NOT a demonstrated
visible-label latency improvement. The performance result is VCS completion.
Draft-page duplicate selector and create-branch paths were not measured.

Analysis: analyze.cjs reads captured NDJSON span timing fields and run JSONL.
sum.cjs is the reporter's original git-span summarizer.
Sanitized trace files retain span names, timing, ids and ancestry only;
attributes/events/exit payloads were removed for privacy by sanitize-trace.py.
Before/after driver transcripts preserve request sequences; runner directory
paths were normalized to C:/p/perf for privacy. No timing values were changed.
Video environment label is masked; branch controls and timing are unmodified.
No pairing credential is in these files. All application data is synthetic.

Reproduction scripts: setup.ps1, run-side.ps1, measure.ps1, common.cjs,
branch.cjs. Place driver scripts in C:/p/perf; execute the PowerShell scripts.
Wait for the real server to listen before measure.ps1. setup.ps1 installs and
builds both worktrees and creates the shared repository; Git must already have
the two commits checked out at C:/p/base and C:/p/head. run-side.ps1 takes
base/head. measure.ps1 takes base/head. Stop the captured server process tree
before starting the other side. Each result has a timestamped directory.

The initial wrapper stopped on a Vite+ stderr notice; resumed at the builds
with native exit-code checks. The first seed attempt preceded server readiness;
no switches were measured until project and thread creation succeeded.
