RAG: Add citation tracking and attribution system

[sfloess]

Part of #2: Add RAG (Retrieval Augmented Generation) with citations

Scope

Implement citation tracking that maps generated answer segments back to their source chunks, providing transparent attribution and verifiability.

Details

• Parse LLM output to extract inline citation references
• Map citations back to original source documents with metadata (filename, chunk index, similarity score)
• Produce a structured citations list alongside the answer
• Support citation formats: inline [Source: X] and appended bibliography
• Track citation coverage: what percentage of the answer is grounded in retrieved sources

Output Structure

class RAGResult:
    answer: str                    # Generated answer text
    citations: List[Citation]      # Source attributions
    retrieval_scores: List[float]  # Per-chunk relevance scores
    coverage: float                # % of answer grounded in sources

class Citation:
    source: str        # e.g., "consensus-guide.md"
    chunk_index: int   # Which chunk in the source
    similarity: float  # Retrieval similarity score
    text_span: str     # The cited portion of the answer

Acceptance Criteria

• Citations are extracted from LLM output and mapped to source chunks
• Structured citation objects include source, score, and text span
• Coverage metric indicates how well-grounded the answer is
• Citations are deduplicated (same source referenced multiple times)
• Unit tests for citation parsing, mapping, and coverage calculation