# Reproduce the checkpoint timing comparison

Use Linux x86_64, Git 2.43.0, Node 26.13.1, 8 CPUs; install the checkout dependencies with `vp i`. Prepare separate clean checkouts of [ddb7fb3f50](https://github.com/pingdotgg/t3code/commit/ddb7fb3f50c5c8ddb6d495e3cee04ac677ca589e) and [c302d9f64c](https://github.com/pingdotgg/t3code/commit/c302d9f64c404a7fbeea19f68082efe35a994ba7).

Download `prove.py` and `render.py`. Rendering needs Python Pillow and Matplotlib, DejaVu Sans Mono, and ffmpeg. The driver needs Python, Node and Git. Run:

```sh
python3 prove.py --base-checkout BASE_CHECKOUT --head-checkout HEAD_CHECKOUT --output NEW_RUN_DIRECTORY
python3 render.py NEW_RUN_DIRECTORY
```

The output directory must not already exist. The driver makes one deterministic synthetic repository (5,000 committed text files, 2,000 unique pseudorandom 1 MiB untracked files), copies it for each side, and primes each Git object store using a separate index. This avoids measuring initial compression; the files remain untracked in the real index. Each side starts its own real development server and fresh T3 home with dev seeding disabled. Paid providers are disabled. The real CLI drives devReplay/showcase turns through the provider adapter and orchestration.

There is one first turn, six sequential measured turns, then three batches of four concurrent threads in the same repository. Span names are GitVcsDriver.checkpoints.captureCheckpoint, orchestrationV2.checkpointCapture.execute and RunFinalizationService.finalize. Trace rotations are read oldest first. Successful span timings and interrupted baseline captures are reported separately. CLI wall time includes CLI startup and waiting for completion. No timing threshold affects the workload.

After timing, every created checkpoint tree is compared with independent Git blob hashes for all 7,000 files. The real index must remain unchanged. A further turn checks a same-size untracked edit, addition/deletion, a tracked edit and an unchanged control file. Both commits produce identical unchanged-workspace trees and pass the changed-file checks.

The published comparison uses run-05 only. Its first turn is separate (base 16.818 s, head 14.423 s). The first measured head baseline still takes 3.331 s and is retained in the samples. Cold cache population is not eliminated. Baseline timeout counts across the 19 measured/first turns are five on base and one on head; all final captures and finalizations succeed. That yields 33 versus 37 unchanged-workspace refs; both sides also create two correct changed-file refs.

The MP4s are timestamp-faithful renderings of captured CLI events at 1x, not web UI screencasts. The waiting periods are preserved. The PNG charts share axes. measurement.json contains exact successful/aborted spans, raw wall times, correctness results and summary statistics.

The redacted diagnostic archive also retains preliminary attempts: run-01 encountered cold compression and was interrupted; run-02 was interrupted while adapting baseline-timeout handling; run-03 exposed trace rotation in the initial reader; run-04 completed but capture-tool dependencies were installed during part of its base concurrency measurement. Run-05 repeated both sides with those dependencies already installed and no such work during measurement. Only run-05 is used for the verdict. Credentials, private paths, machine labels, span attributes and exception stacks are omitted from the archive.
