Windows idle remote-status proof

Environment: same leased Windows Server 2025 machine for both commits; Git 2.56.1.windows.1; Node 26.13.1; headless Chromium, viewport 1600x1000, dark theme.
Base: dcd8df806daff06129963cc2ab011897035e18f3
Head: 7b90485a0fec4b72bec569c39e4d19e2f0134ad9

Build each checkout: vp i; then from apps/web run vp build.
prepare.ps1 creates two local-clone repositories from the checkout, with feature-x one commit ahead of main. p0 has a GitHub origin and tracking main; p1 has no remote.
seed-side.ps1 starts the real server with T3CODE_DEV_SEED=0 and --dev-url solely to enable the replay provider, adds projects through the CLI, and completes one todo-list replay thread per project. It restarts without --dev-url to serve that checkout's built web app, waits for readiness, and mints a pairing link. Pairing links are private inputs and not included in these artifacts.
run-side.ps1 runs seeding and measure.cjs and stops the server. Each side has a fresh T3 home. Paths in PowerShell scripts use USERPROFILE in place of the private working-directory prefix used during capture; the commands and measurements are otherwise the same.

measure.cjs arguments: playwright module path, private pairing-output file, new output directory, synthetic repository directory. It opens the real app, waits for thread rows, clicks each thread, settles 15 seconds, records three 95-second idle windows, captures a stable screenshot, checks out feature-y, waits 35 seconds, captures it, and checks out feature-x. It records video and selected WebSocket status fields. No API polling substitutes for browser-driven status subscriptions.

analyze.cjs arguments: comma-separated trace files (include rotated files if present), measurement directory, before/after label, commit SHA. It counts runGitCommand descendants and wall durations of VcsStatusBroadcaster.readRemoteStatus spans starting inside each idle window. It reports every poll, operation attribution, per-window medians, pooled medians, and max-minus-min window-median spread. These are real app/server measurements, not test-run durations.

Raw traces contain absolute paths, so the public captures are the analyzer's measurement output, not the raw server log. No spans are filtered by duration or process count. The video omits its first 45 seconds to exclude pairing/startup (this may also omit part of the first idle interval; the measurement output retains every measured poll) and masks only the machine display label. Screenshots mask the same label. No playback speed change is applied.

The requested base already includes the separate base-branch-resolution cache optimization. Compare the supplied base/head measurements rather than the higher counts from the original report's older revision.

After an SSH interruption stopped an incomplete head capture, resume-head.ps1 reset feature-x, restarted that same fresh head home, and reran the complete measurement through a background Windows process with output redirected to a file. Only the complete rerun is used for the final head comparison. The server and driver were stopped after capture.

A complete base confirmation was then recorded after the head, with feature-y already present at the same commit as feature-x on both synthetic repositories, matching the head rerun's starting branch state. The final comparison uses this later base confirmation; the initial base measurement is retained as corroboration. Both complete captures use the same measure.cjs driver. To recreate that final starting state from fresh fixture repositories, create feature-y at feature-x before running each side.

Final timing comparison: base confirmation 444.87/421.63 ms (no remote/GitHub origin), head 368.23/329.57 ms. Nine polls per cell, three 95-second windows. Improvement 76.64/92.06 ms exceeds both sides' window-median spreads (base 6.19/8.08 ms, head 10.42/20.68 ms).

A caveat outside all timing windows: one combined head status snapshot after returning to feature-x reported aheadCount=0 and aheadOfDefaultCount=0; the remote update restored 1 about 1.2 seconds later. It did not recur in the final base confirmation. Causality was not established. Steady idle remote values and final branch/count matched on both sides.

The bundle includes sanitized measurement trace spans (names, identifiers, parents, timestamps; all attributes omitted), and base-data/head-data with the original window boundaries, rows, and selected status events. Decompress base-trace.ndjson.gz and run: node analyze.cjs base-trace.ndjson base-data before dcd8df806daff06129963cc2ab011897035e18f3. Use the head files and head SHA for the other side. trace-sanitize.cjs documents the sanitization; it retains the measured readRemoteStatus spans and their descendants, with no duration/process-count filtering.
