Windows real-app performance proof Prerequisites: the supplied base and head commits checked out at C:/p/base and C:/p/head, each installed with `vp i`. The same Windows pool machine ran all six windows. No provider credentials or live user database were used. 1. Run prepare.ps1. It builds both web clients, pushes the base snapshot into a new local bare origin, makes a non-local clone, and pushes feature with -u. 2. Run stamp.cjs with Node to verify the no-op fetch directory timestamp effect. 3. Run run.cmd. It resolves the actual Node executable before starting measure.cjs (process.execPath through the Vite+ launcher can otherwise be the launcher). 4. The driver starts each commit's source server with its own fresh T3 home and built web assets. It adds the same repository through the CLI, starts a free Replay command-output thread, pairs Chromium, and opens the thread by its environment/thread route. It waits for Git action options plus 15 seconds, then records 180 seconds idle. It changes neither refs nor app polling rates. 5. Order: base1, head1, head2, base2, base3, head3. Only one server runs at a time. Each server is stopped by its captured PID. Every sample comes from the same repository, browser version, viewport (1440x900), and dark theme. 6. The server's real trace spans are counted by parentSpanId ancestry. Git process spans are named VcsProcess.spawn or runGitCommand. The uncached resolution span is resolveBaseBranchForNoUpstream; its cached wrapper is excluded. 7. Measurements are written to a fresh timestamp directory per invocation. Run `python report.py ` to reproduce the text report and both PNG captures from measurements.json. 8. The representative recordings were converted with: ffmpeg -ss 20 -i before.webm -t 180 -c:v libx264 -preset veryfast -crf 28 -pix_fmt yuv420p -an before-idle.mp4 ffmpeg -ss 20 -i after.webm -t 180 -c:v libx264 -preset veryfast -crf 28 -pix_fmt yuv420p -an after-idle.mp4 This omits the opening load screen; the measurement JSON retains each full 180-second window, irrespective of the recording excerpt. Scope: natural background remote status polls with Git action options closed. The separately reported open Git-actions-menu workload was not measured. The final base/head pair was run with resume-final-pair.cmd after one fresh base server exited during navigation, before its timed window. The four prior complete windows were retained. The final measurements.json concatenates the two successful invocation lists, in the recorded order. export-traces.cjs exports the measured poll span trees from each invocation; the combined gzip contains the concatenation, with attributes and events removed for privacy. The primary timing statistic is the median of the three per-window medians. Between-run spread is the range of those per-window medians. The text/JSON reports also include the pooled median and every individual poll duration.