llm-judge passes on unclamped scores, a 0-100 reply always passes

[royalpinto007]

Problem

A stub judge returning {"score": 85} on a wrong answer yields {score: 1, passed: true}: the raw number is compared to the threshold before clamping.

Where

src/scorers/llm-judge.ts (Number(obj.score), passed = score >= threshold; util.ts clamps only the reported score).

Acceptance

Add a failing test first asserting score 85 yields passed: false/parse-reject. Then clamp (or reject) before the pass check.