Issue 40247 — manual Windows app proof

Base: 1425613b9d47342aa78a9005843c3db0c9cd2186
Head: ec47b4fe02be277798ad99f76bc923b0f5aa4525
Same leased Windows Server 2025 machine, Defender real-time protection on.
Git 2.56.1.windows.1, Node 26.13.1, Playwright 1.64.1 Chromium, 1600x1000 light theme.

Create C:/proof-base and C:/proof-head worktrees at the respective commits.
Run vp i and vp build from apps/web in each. setup.ps1 contains the head build and
shared fixture seed; base-setup.ps1 contains the base build. The source clone path
is expressed portably as USERPROFILE/t3code in these published scripts.
The shared fixture has 32169 tracked files, a 4501827-byte index, 40 tracked .ts
edits, and 400 new .ts files (8 newmod directories of 50 files each).

Run run-side.ps1 base, then run-side.ps1 head, allowing all setup/build processes
to finish before measurement. Each side seeds its own fresh server home through
project add and thread start with devReplay/todo-list, then restarts the server
without --dev-url to serve its own production build. The same diff.cjs drives both.
It pairs, opens task 4-0, presses Ctrl+D, wheels 2500 px every 250 ms for 25 s,
and repeats with a fresh page three times. Timings use WebSocket send to successful
response. Per-file requests interrupted before a successful reply are excluded.
Videos are representative loads cropped to the Diff panel, beginning 15 seconds into the recording
for privacy; they are supplementary, not the timing source.

base-public.json and head-public.json retain captured responses and timings;
only cwd fields were replaced with <fixture>. To rerun analyze.py, name copies
base.json/head.json. It compares exact sources, including patch text and file
metadata, excluding the envelope cwd and generatedAt timestamp.
spans.py reads the server trace files collected by run-side.ps1. The uploaded
spans-summary.json contains only derived timings and Git operation names; raw
server traces are not published because they also contain machine-local data.
matched-files.json compares only paths successfully returned in all six loads.
The proof did not use tests or a synthetic Git microbenchmark.
