Performance proof: no-remote Git status refresh

Environment: Windows Server 2025 Datacenter 10.0.26100, Intel Xeon Platinum
8370C, 4 logical CPUs, 16 GB RAM; Git 2.56.1.windows.1; Node 26.13.1.

Base: d51d8c4766f6c8760d5db9428596b98c4f81eab1
Head: dbee98076981b1d638b549fb72f198aa027177ab

Start each installed checkout with start.ps1 -Checkout <checkout> -HomeDir
<fresh-home> -Port <unique-server-port>. Uses T3CODE_DEV_SEED=0 and a fresh app
home. Wait for server-runtime.json and its startup pairing announcement.

Run node measure.cjs <checkout> <home> <new-output-directory> before (or after).
It clones the checkout locally, removes origin, creates feature-5 and an empty
commit, adds the project through the real CLI, starts Replay (dev)/showcase,
and opens its thread in headless Chromium (1440x1000, dark).

After both apps/browsers have warmed, create <base-home>/measure.ready. Wait
for that run to finish, then create <head-home>/measure.ready. Timed loops do
not overlap. Each run creates 33 empty commits, waits 1500 ms after each,
times the real vcs.refreshStatus WebSocket RPC, and waits another 1000 ms.
The first three timings are warmups; five consecutive batches contain six
measured refreshes each. These are forced full refreshes, not just the 30-s
remote poll. Both browser subscriptions also exercised ordinary polling.

measure.cjs saves samples, complete returned status, a final screenshot,
and the browser video. Its final remote-add check is outside the timings.
extract.cjs <home> <output> <side> selects git spans by synthetic cwd,
retains their ancestor tree, and strips paths/attributes. Git durations are
cumulative subprocess spans and may overlap; RPC duration is wall time.

report.py <before-raw.json> <after-raw.json> <output> computes conventional
medians from all samples and six-sample batches, normalizes only the synthetic
workingTree.rootPath, verifies no remote/provider/PR in every sample, and
renders the measurement PNGs. Raw response paths are the only redaction.

probe-remote.cjs <head-checkout> <head-home> <output-json> changes the existing
origin to the public GitHub repository URL, invokes the next localOnly refresh
(no fetch), captures provider detection, and restores the local URL.

capture.cjs <checkout> <home> <output> <side> removes the added origin, opens
an authenticated browser on the existing thread, waits for Replay completion
and Publish repository, and accepts a still only after two identical PNGs
500 ms apart. A previously positive remote cache can delay this post-check
recapture; it does not affect the original no-remote measurement.

The paired videos are trimmed from each browser recording to the interval
from the first timed request through the last response. No speed change.

After the remote-add/provider probe, the head server was restarted to clear
its existing positive-remote cache for the final no-remote still. All timing
samples and measurement recordings precede that restart.
