[p-r-a-v-i-n]

CodexHarness.parse now removes OpenAI citation markers from a Codex result. Before, a Codex run that searched the web could leave markers in its result: U+E200 cite, a U+E202 reference ID, and U+E201. They showed as stray text in saved results and in the prompts of later runs that read those results.

• parse serializes the payload with ensure_ascii=False, removes each marker and the spaces before it with _CITATION_MARKER, and loads the JSON again. One pass covers every nested string.
• The pattern stops at a quote, so a match stays inside one JSON string. A marker with no closing U+E201 is left in place, and the JSON stays valid.
• output.json is written from the cleaned payload.

[czpython]

The round trip in parse is correct. Checked with quotes, backslashes, newlines, nested lists, and a marker written as a \ue200 escape. Each one comes back clean and valid.

The pattern needs two changes. See the inline comments. The test needs the same escapes.