Windows command-discovery responsiveness proof

Commits:
  base: 9f19a6ddf57d5d764c9d36eb58f07a4bb196ff05
  head: 7fbdb68b41f58fb4894755a3f9f84b61c7fc69a0

Environment: Windows Server 2025 Datacenter, 10.0.26100 x64; Node 26.13.1;
Playwright Chromium headless shell build 1248; dark theme, viewport 1440x1000.
Both sides run sequentially on the same Windows machine, with vp-installed
worktrees and separate synthetic T3 homes. No real model turn is run.

Setup (parameters are paths chosen by the operator):
  powershell -File launch.ps1 <checkout> <scratch> <base|head>
  powershell -File seed-portable.ps1 <checkout> <scratch>
The launch is started via Win32_Process.Create to survive a normal SSH return;
retain its returned process ID and stop that process tree with taskkill /T /F.
Wait for the server ready announcement before seeding or driving the browser.
The only seed is an empty git project titled "Command discovery demo".
A new home is used for each commit, never a copy of live state.

Run the ordinary-PATH functional control once before each side's cold trials:
  node control.mjs <checkout> <scratch> <base|head>
Then run the same driver on each commit:
  node drive.mjs <checkout> <scratch> <base|head> 3
  node drive.mjs <checkout> <scratch> <base|head> 2 10
Each side uses the same two browser batches, three plus two trials. The optional
last argument offsets the fresh UNC host number, so the second batch never
reuses a failed host. combine.py joins the two JSON files and preserves batch
and original round numbers. The driver records its process ID, JSON output,
screenshots and a browser video.
Do not interrupt SSH to stop a Windows driver: explicitly stop its recorded PID.

For each cold trial, instrument.mjs adds an unreachable UNC share first on PATH
inside the running server. Each trial uses a fresh 10.255.x.y host, avoiding the
Windows failed-host cache and the vp Node shim's startup PATH walk. PATHEXT is
.COM;.EXE;.BAT;.CMD, and uvx is absent. The preload leaves the resolver unchanged.
It measures the longest interval between 5 ms timer callbacks in that server.
Its local control endpoint restores the original PATH after every trial.

The driver opens Settings > Providers, clicks Add provider and Show more, and
types gemini in Search ACP Registry. It waits for the Add provider dialog's
compatible-agent status and Gemini result, for up to 120 seconds. A separate
Node HTTP client issues a request every 50 ms to /api/evidence-heartbeat through
Vite's /api proxy. This intentionally missing URL returns a real server HTTP
404; the measured response time shows whether other server callbacks can run.
The driver stores the slowest response per trial and the server's longest gap.
Search timeouts on base are preserved as failures, not omitted from the median.

Read base.json and head.json for every recorded trial and the returned UI text.
The two functional-control JSON files record the ordinary-PATH result text.
render.py renders the measured numbers into the paired images and summary.json:
  python render.py base.json head.json <output-directory>
The chart images are a rendering of the JSON, not screenshots of an assertion.
Browser recordings show the actual interaction. Their initial pairing/startup
prefix is trimmed; the measured interactions remain at real speed. Script comments and portable
launch/seed wrappers explain setup; no production data or credentials are used.

Video processing: remove the first 9 seconds of each browser batch, transcode
to H.264 MP4, and concatenate the three-trial and two-trial recordings. No
measured interaction is sped up. The original JSON times are authoritative.

Prepared by gpt-6.1-sol using the Codex harness.
