# Real-app Windows project search proof

No mocks and no test-suite results are used as evidence. Both versions run `vp run dev` with isolated empty T3 homes and `T3CODE_DEV_SEED=0`. Projects and replay threads are created through the app CLI. Playwright drives the real Chromium client locally on Windows.

App revisions:
- Base: f0a33213547ce81ea9c39e4bedb13829cf307f56
- Fix: 3e7f258c343af4e25fb568dbd065373c2f59e5ca

Each run creates a fresh `git clone --no-hardlinks` of the base file tree (32,347 tracked files) outside Defender exclusions. Both sides therefore search identical files. The browser uses a 1440x1000 viewport and renders the dark theme. The startup/authentication, clone and replay turn occur before the timer starts. The measured action is Ctrl+Shift+F, fill `waitForIndexReady`, then wait for the result status to appear. Time includes the UI debounce and result rendering. A timer overlay annotates the recording without changing search data. Final cold screenshots wait for matching screenshots 500 ms apart.

Three cold runs per app revision, sequentially on the same Windows Server 2025 machine. The driver waits 210 seconds between runs to limit overlapping background index activity. Failed base searches keep the dialog open and refill the same query every 33 seconds until results appear. The base recording therefore also measures the first successful answer. On head the initial response is already successful, so the initial response and first success are the same measurement.

`head-driver.cjs` and `base-driver.cjs` are the exact drivers used. Their initial search and timer are the same. The base driver adds the continued recording/retries after an error to capture eventual successful results. The original head driver's metadata says `theme: light`; this was a metadata typo (all three head screenshots and videos render dark). The base metadata and comparison report state the rendered dark theme.

To repeat on an equivalent leased Windows machine, install both app revisions in `C:\bench\proof-base` and `C:\bench\proof-head` with `vp i`. Set `T3_PROOF_SOURCE` to an installed repository containing `scripts/node_modules/playwright-core`. Install the Chromium headless shell version referenced by the drivers. Start each side with `powershell -File start.ps1 base` / `head`, wait for startup, then run `node head-driver.cjs head` and `node base-driver.cjs base` sequentially. Keep every cold-clone directory outside exclusions. `start.ps1` records each process-tree root PID for cleanup. End those process trees after the run.

Raw output is in `base.jsonl` and `head.jsonl`. `analyze.py base.jsonl head.jsonl <report-directory>` creates the comparison JSON and HTML. It compares all successful result rows after sorting the file groups, because the native index and git grep return groups in different orders. The result sets contain the same 16 matching lines in the same 3 files.

Published app images and recordings are cropped from the original 1440x1000 viewport to the search dialog and benchmark caption (640x650, x=400, y=70), to remove machine details. Crop command: `ffmpeg -i input.webm -vf crop=640:650:400:70 -c:v libx264 -preset fast -crf 22 output.mp4`. No timing changes or speedups are applied. Stills use the identical crop.

Fresh clones stand in for Defender definition cache invalidation. A definition update and WSL were not run. This proves the reported cold first-search behavior on Windows Server, not Windows 11-specific behavior or all query option combinations.
