Windows large-diff proof driver

Prerequisites: Git, Node, vp, the repository dependencies, a production web build in each checkout, and Playwright Chromium/ffmpeg installed. Leave Defender enabled.

Build each commit separately with vp i, then vp build from apps/web.
Create one shared synthetic repository:
  node seed-repository.cjs <base-checkout> <scratch>/repos/p0 d51d8c4766f6c8760d5db9428596b98c4f81eab1
Keep seed.cjs, diff.cjs, and run.cjs together in <scratch>.
Run sequentially on the same Windows machine:
  node run.cjs before <base-checkout> <scratch>
  node run.cjs after <head-checkout> <scratch>
The runner uses one fresh home per invocation, registers the replay provider while seeding three todo-list threads, then restarts without --dev-url to serve the production build.
The browser driver pairs, opens thread 0-0, closes any open diff, presses Ctrl+D, and scrolls 2500 pixels every 250 ms for 60 seconds. Three independent page loads use the same 1600x1000 viewport. It records websocket request/response timing and returned patch data, screenshots, and video.

After copying the generated output to the analysis machine:
  python3 summarize.py <before-output-directory>
  python3 summarize.py <after-output-directory>
The summarizer reads measure.json plus traces/server.trace.ndjson* and writes summary.json and contents.json. Content comparison checks source kind, diffHash and file statistics, excluding generated timestamps and revision metadata.
Place these outputs under before/ and after/ to render the measurement captures:
  python3 render-report.py <comparison-directory>

Measurement definitions:
- Client medians use successfully completed single-file review.getDiffPreview replies. The JSON summaries separately list sent requests. Outstanding requests at page close are excluded from latency medians.
- Unique files count successfully returned patches. More files return on head during the fixed 60-second interaction.
- Server git-request counts cover the complete three-page-load trace, including work started before the client's per-load recording window. They therefore differ slightly from client reply counts.
- requests.csv contains the completed client timing samples; git-requests.csv contains each qualifying server request's git operations and durations. Local repository paths and credentials are omitted.
- compare.py checks identical diffHash and per-file statistics for the intersection of returned paths, and computes medians restricted to that intersection. Whole-preview responses stream and are not included in this content comparison.
- The representative MP4 for each side is the final 65 seconds of load 2, retaining the scroll and settle phase while omitting setup screens.
