Make grading summary deterministic and JSON-safe across report/finalize flows

[Copilot]

The current summary field was derived from LLM-generated prose, which made output non-deterministic and inconsistent with strict JSON-first consumers. This change makes the summary fully deterministic from scoring data and enforces a plain-text, JSON-safe summary shape everywhere it is produced.

• Deterministic summary contract

  • Added build_deterministic_summary(...) in src/mas/tools/summary_builder.py.
  • Summary is now computed from total_score, total_marks, grade, and per-criterion scores (not from model prose).
  • Output format is stable and single-line, with explicit score breakdown.

• Replaced LLM-paragraph extraction

  • Removed “first prose paragraph” summary extraction logic from:
    • src/mas/agents/report.py
    • src/mas/agents/finalize.py
  • Both agents now call the shared deterministic builder, ensuring identical summary semantics in sync and finalize paths.

• Behavioral coverage updates

  • Updated report/finalize tests to assert deterministic summary output instead of markdown paragraph heuristics.
  • Added dedicated unit tests in tests/test_summary_builder.py for:
    • full criterion breakdown output
    • empty-criteria fallback behavior

• Docs alignment

  • Updated README API response example to reflect deterministic summary format.

summary = build_deterministic_summary(
    scored_criteria=[
        {"name": "Definition of Containerisation", "score": 0, "max_score": 5},
        {"name": "Benefits of Containerisation", "score": 3, "max_score": 5},
        {"name": "Role of Kubernetes", "score": 1, "max_score": 5},
        {"name": "Technical Accuracy and Depth", "score": 2, "max_score": 5},
    ],
    total_score=6,
    total_marks=20,
    grade="F",
)
# "Total score: 6/20 (30.00%), grade F. Breakdown: Definition of Containerisation: 0/5; Benefits of Containerisation: 3/5; Role of Kubernetes: 1/5; Technical Accuracy and Depth: 2/5."