fix: parse judge replies that prefix prose before the fenced block

[Mr-Neutr0n]

Problem

_parse_judge_json required the reply to start with a backtick fence:

if cleaned.startswith("```"):
    cleaned = cleaned.split("```")[1]

Judge models frequently preface the block with a sentence (Certainly. followed by a json fence). For those replies the guard is False, the raw text hits json.loads, parsing returns None, and _judge_one scores the criterion as JUDGE_PARSE_FAILED (unmet) after burning all retries on a perfectly valid judgment.

That silently deflates HealthBench / HealthBench Professional scores for any judge model that wraps its JSON in conversational text, and it looks like a model-quality gap rather than a parser gap.

Change

• Find the first fence anywhere in the reply, strip an optional json tag, and parse up to the closing fence.
• When there is no fence at all, fall back to the outermost {...} object so prose-wrapped judgments still parse.
• Non-bool criteria_met, malformed JSON, and no-object replies still return None exactly as before.

Verification

pytest tests/test_judge.py -q

Same pass/fail set as before the change plus the new test_parse_judge_json_handles_prose_around_the_block (5 cases). The 3 pre-existing failures in that file reproduce identically on clean main without optional provider deps installed; none touch the parser.

---

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