Issue 40183 real-app performance proof

Platform: Windows Server 2025 Datacenter, 4 logical CPUs; Chromium 156 headless;
built web client; 1440x900 viewport; dark theme; devReplay todo-list provider.

Base: dcd8df806daff06129963cc2ab011897035e18f3
Head: 1d734f6570b39a0d097bdf44e27c8b4ec3e3d61c
Common fixture-source checkout: ca6e09c3e1191c8d46a9b97dd658ccb65dde8fcb

Install each checkout with vp i, then run vp build from its apps/web directory.
On Windows, execute the same measure.cjs driver for each checkout:
  node measure.cjs BASE_CHECKOUT BASE_OUTPUT FIXTURE_SOURCE_CHECKOUT
  node measure.cjs HEAD_CHECKOUT HEAD_OUTPUT FIXTURE_SOURCE_CHECKOUT "BASE_OUTPUT\Changes workload"
Use new output directories. Each side initializes its own synthetic server database.
The optional fourth argument reuses exactly the same git workload. The driver can
resume a seeded server using EVIDENCE_HOME and EVIDENCE_THREAD_ID; this was used
for the final base capture after correcting the driver, without changing the workload.
All server processes start with T3CODE_DEV_SEED=0 and end in the driver's finally.

The fixture is a local clone with origin removed, feature-3 branched from main,
one committed feature.txt, and a blank line plus // edit appended to the first
800 tracked *.ts paths. It has 800 modified tracked files and zero untracked.
Changes vs main includes the new committed file: 801 files, +1601/-0.

The driver pairs Chromium, opens the replay thread, and repeats 16 times:
reload thread, wait for the replay message, wait 3 seconds, Ctrl+D, wait at least
7 seconds and until the untracked check finishes, capture on cycles 0 and 15,
close the Diff tab using Close Diff, then wait 500 ms. Still captures require
two identical PNGs 500 ms apart. Cycle 0 is excluded as warmup.

The paired tracked/complete totals requests start during thread navigation.
Complete-only refreshes caused by opening the panel are not counted. The driver
labels phases and rejoins T3's split WebSocket frames before decoding responses.
Analysis measures each complete response and compares the full per-file stats.
Server spans supply git operation counts and server durations; WebSocket packet
arrival times supply full-pair elapsed time. Browser spans align server requests
by start time. No tests are used as the performance evidence.

Recreate the reports from the sanitized measurement inputs:
  python analyze.py base-input-requests.json base-input-traces.ndjson base-results.json
  python analyze.py head-input-requests.json head-input-traces.ndjson head-results.json
  python report.py .
  convert timings.svg timings.png

The sanitization removes host paths, machine labels, credentials, resource metadata
and error stacks, while retaining IDs required to correlate durations and synthetic
file statistics. sanitize.py records that transformation. Raw private logs are
not published. The videos show the last 23 seconds of the same scripted flow;
only the environment-label rectangle is redacted, identically on both sides.

Scope: the reported no-untracked workload and paired preview path. Nonempty
untracked trees, token expiry/fallback, other operating systems and other clients
were not measured. The unrelated complete-only panel refresh is not claimed faster.
