Add CAR evaluation metadata and improve structured answer parsing

[ChathurangiShyalika]

Summary

• Added support for per-scenario groundtruth_eval.json metadata.
• Updated CAR scoring to check mode key, one-key format, and required term coverage.
• Added CAR soft score and aggregate reporting fields.
• Fixed structured answer parsing so final JSON is preferred over earlier parenthesized prose.
• Added regression tests for CAR metadata loading, CAR scoring, and noisy JSON parsing.

Validation

.venv/bin/python -m pytest src/evaluation/tests/test_static_json_scorer.py src/evaluation/tests/test_loader.py