fix: parse structured output wrapped in a markdown code fence

[andyli953]

Problem

parseStructuredOutput in plugins/codex/scripts/lib/codex.mjs calls JSON.parse on the raw final message. A model asked to reply with JSON frequently wraps it in a fence, so the parse throws:

Codex did not return valid structured JSON.
- Parse error: Unexpected token '`', "```json
{"... is not valid JSON

The review then falls back to printing the raw message. Nothing is lost, but the structured path — verdict, findings, per-finding file and line — is unavailable.

I hit this on 2 of 3 consecutive /codex:adversarial-review runs against the same repository. Identical prompts; whether the reply arrived fenced or bare varied run to run, so it is not deterministic and not something the caller can avoid.

Fix

Strip a fence, but only one that encloses the entire message:

const FENCED_MESSAGE = /^```[A-Za-z0-9_+-]*[ \t]*\r?\n([\s\S]*?)\r?\n?```[ \t]*$/;

That keeps JSON carrying backticks inside string values intact, and leaves malformed output, prose and empty messages failing exactly as they do today. This is deliberately narrow: it only recovers output that never reached the schema check. The existing graceful degradation for an unexpected review shape (renderReviewResult, covered in tests/render.test.mjs) is untouched.

Tests

New tests/codex-structured-output.test.mjs, in the existing node:test style.

	pristine	with fix
bare JSON	pass	pass
```json fenced	fail	pass
fence with no language tag	fail	pass
fenced with CRLF	fail	pass
backticks inside a string value	fail	pass
malformed JSON	pass (rejected)	pass (rejected)
prose, not JSON	pass (rejected)	pass (rejected)
empty message	pass	pass

Four cases fail against the current parser and pass with the change; the other four pass either way, so the fix cannot be silently loosening the parser into one that accepts anything.

Full suite

npm test on macOS / Node v23.11.0, before and after:

• before: 87 passing, 4 failing
• after: 95 passing (+8, exactly the new tests), same 4 failing

The failing set is byte-identical before and after:

resolveStateDir uses a temp-backed per-workspace directory
result returns the stored output for the latest finished job by default
status preserves adversarial review kind labels
status shows phases, hints, and the latest finished job

These four fail on pristine main in my environment and are unrelated to this change — I have not investigated them and am not claiming they are broken upstream, only that this PR neither causes nor fixes them.