What changed

Keep one snapshot of submitted image File objects and names for each playground run. Use it for multipart upload and success/error history entries. Mark current tiles as scored only when they still refer to those submitted files; restoring history retains the original image files.

Verification

Real Chromium and an owned local HTTP server confirm the actual multipart body before allowing the response to finish. Upload image A, replace it with image B during the response wait, then restore the saved history entry.

• Main: three failing cases — success with replaced image, HTTP failure with replaced image, and a text-only request followed by adding an image. History binds the old request to the new files; successful image replacement also incorrectly marks B as scored.
• This branch: all four browser controls pass, including an unchanged-image success control. History restores A or no image as submitted; B is not falsely marked scored. No page errors.
• Full pytest: 713 passed, 136 skipped, four failures, matching unchanged main. Two missing optional mlx_vlm imports and two missing legacy test_p3_labeltrain fixture imports remain.

The native browser protocol fixture runs separately; no model inference or paid API request is claimed. No browser dependency was added to the Python package. Verification used macOS arm64, Chromium and Python 3.13.5 with offline model settings and an empty owned Hugging Face cache. No Windows/browser matrix or hosted CI pass is claimed.

Prepared with Codex assistance. Signed commit and Rudy Celekli DCO trailer included.