# Authored reference repair — not evaluated client output

The original source and failing ZIP are unchanged. `manual-repair/` is a separate authored correction, not a response from a Claude Code/Codex session. No coding client or paid model was run.

Run `python3 verify-manual-repair.py`. It executes the original and repaired copies using identical acceptance tests, records stdout/stderr and real exit codes, captures a unified diff, and verifies SHA-256 for the original code/tests/ZIP. Absolute local path prefixes in the saved traceback are replaced by `<example-directory>`; assertion messages are unchanged.

Original: 2 tests with empty-list IndexError. Repair: initialize an empty mapping, 2 tests pass.

The evidence is limited to the supplied task and provided tests; it does not establish malformed-field handling, a live client integration or model quality. `render-video.py --output DIR` renders a disclosed 18-second local reference demo from real test evidence; it is not a client screenshot or terminal-session recording. Full playback/listening and page operation are unverified.
