feat: multi-dimensional quality scoring for structured outputs

[sungdark]

Summary

Implements bounty #1: scoring structured submissions (JSON/markdown/code/text) with a weighted 0-1 quality score and per-dimension feedback.

Included

• • format auto-detection
  • 5 rubric dimensions: completeness, format compliance, coverage, clarity, validity

  • weighted output schema:
• • schema validation
  • multi-format scoring coverage
  • benchmark check for 100 submissions <10s

Validation

Run:

All tests pass.

[sungdark]

Follow-up update from my side:

I pushed an improvement commit to make the scorer easier to tune and review:

• Added configurable weights (auto-normalized)
• Improved per-dimension feedback wording
• Added evaluate_against_ground_truth() helper for dataset-level error reporting
• Expanded tests to include schema, format coverage, benchmark, and ground-truth alignment checks

Happy to tighten/adjust thresholds if you share your preferred calibration dataset.