fix: grade decorated Answer markers instead of falling through to prose

[Mr-Neutr0n]

Problem

The MCQ answer-marker regex only allowed whitespace between ANSWER/IS and the letter:

(?:FINAL\s+)?ANSWER\s*(?:IS\s+)?[:=]?\s*\(?([A-J])\)?\b

Real model replies decorate the marker, and the most common form in the wild is markdown bold: **Answer:** C. That fails the marker rule, so grading falls through to the standalone last-occurrence fallback, which happily takes a letter from later prose. Concretely, with gold C:

**Answer:** C. Option D was considered and rejected.

grades as D (wrong). Same for **Final Answer:** A followed by a differential listing (B), (C), etc.

Reproduced before/after across a battery of decorated markers (**Answer:** X, Final Answer — X, fullwidth colon Answer：X, ANSWER IS :: X): old regex missed every one; new regex catches all.

Change

Allow up to six non-alphanumeric characters between the marker words and the letter:

(?:FINAL\s+)?ANSWER\s*(?:IS\s+)?[^A-Za-z0-9]{0,6}\(?([A-J])\s*\)?\b

Adversarial negatives verified unchanged: ANSWERED, digits, out-of-range letters, prose without markers all behave exactly as before. Plain forms (Answer: B, FINAL ANSWER: D, answer is (E)) unchanged. Boxed answers still win.

Verification

pytest tests/test_mcq.py -q   # 9 passed (5 existing + 1 new regression test with 6 cases)

---

This change was prepared with AI assistance under human direction and review.