fix(v1): prefer Final Judgment over draft boxed answers in parse_judge_choice

[crazywriter1]

Supersedes #2008 — same fix, rebased onto current main after #2204 moved scoring helpers to verifiers/v1/utils/score.py.

Summary

• parse_judge_choice previously unwrapped the last \boxed{...} before looking for verdict markers, so a draft box could wipe a later Final Judgment / Final Answer line and return the wrong choice.
• Search order is now: explicit verdict marker on the full reply → boxed answer (if present) → last choice match.
• If a verdict marker is present but has no choice after it, fall through to boxed / last-match instead of returning None.

Risk

Low. Only flips scores when a reply has both a draft \boxed{...} and a later explicit verdict line (or an empty marker that previously failed). Default yes/no-only reference prompts are mostly unaffected.

Test plan

• Manual: Draft: \boxed{A} + Final Judgment: B → B
• Manual: Final Judgment: (empty) + earlier \boxed{A} → A
• Manual: verdict-marker-only, boxed-only, boxed-over-CoT, last-match, incomplete <think> cases

---

Note

Low Risk
Scoring-only change in judge reply parsing; may flip grades when replies mix draft boxes and later verdict lines, with no auth or runtime impact.

Overview
parse_judge_choice no longer replaces the post-</think> text with a strict \boxed{...} extract before hunting verdict lines, so a draft box cannot override a later explicit verdict.

Parsing order is now: scan the full reply for FINAL JUDGMENT / FINAL ANSWER / FINAL VERDICT (then bare JUDGMENT / VERDICT / ANSWER), last marker wins, and the A/B/C (or custom) choice must sit on that marker’s line via new _choice_on_marker_line. If no marker yields a choice (e.g. empty Final Judgment:), it falls back to strict boxed content when present, else the last choice token in the text.

Also drops the noqa: BLE001 on the broad except in verify_boxed_math_answer (behavior unchanged).

^{Reviewed by Cursor Bugbot for commit 08bc1a8. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Fix parse_judge_choice to prefer Final Judgment markers over draft boxed answers

• Adds a _choice_on_marker_line helper in score.py that restricts choice detection to the same line as a verdict marker, preventing later lines from being matched incorrectly.
• Reworks parse_judge_choice to prefer the last "Final" verdict marker over bare markers, and defers boxed-answer extraction until after marker-based matching so earlier boxed content cannot shadow a later verdict line.
• Falls back to a boxed answer choice, then to the last choice match in the text, if no marker-based choice is found.
• Behavioral Change: choices are no longer accepted from lines following a verdict marker; only the marker's own line is scanned.

^{Macroscope summarized 08bc1a8.}

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

Want reviews to match your repository better? Bugbot Learning can learn team-specific rules from PR activity. A team admin can enable Learning in the Cursor dashboard.

^{Reviewed by Cursor Bugbot for commit c404576. Configure here.}