Add graduated rubric scoring (partial credit)

[ScuttleBot]

Goal

Convert binary 0/1 criteria to graduated 0.0 / 0.25 / 0.5 / 0.75 / 1.0 scales to better differentiate model quality.

Background

Current tasks often score 0 or 1 per criterion, missing subtle differences in solution quality. A model that produces a correct but inefficient solution scores the same as one with an elegant approach.

Implementation

1. Update grading infrastructure to support graduated scores
2. For each criterion, define quality levels:
   • 0.0: Missing/wrong
   • 0.25: Partially correct, major issues
   • 0.5: Mostly correct, some issues
   • 0.75: Correct with minor issues
   • 1.0: Fully correct/excellent
3. Update existing tasks to use graduated criteria where appropriate

Success Criteria

• Score distribution should show more variance (not clustering at 0 and 1)
• Model rankings should be more stable (less sensitivity to binary cutoffs)

References

• RACE Benchmark multi-dimensional scoring
• Item Response Theory discrimination parameters