RAG: Build context assembly from retrieved chunks

[sfloess]

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

Scope

Implement the context builder that takes reranked chunks and assembles them into a structured context window suitable for LLM consumption.

Details

• Accept top-k reranked chunks with metadata (source, score, position)
• Assemble chunks into a coherent context string with source attribution markers
• Respect token budget constraints (configurable max context tokens)
• Handle chunk overlap and deduplication (adjacent chunks from same document)
• Preserve source metadata for downstream citation tracking
• Order chunks by relevance score within the context window

Context Format

[Source: consensus-guide.md, chunk 3, score: 0.92]
Multi-model consensus uses 3+ AI models to validate outputs...

[Source: arbiter-pattern.md, chunk 1, score: 0.87]
The arbiter/worker pattern distributes tasks across workers...

Acceptance Criteria

• Context builder produces structured context from ranked chunks
• Token budget is enforced (truncation with priority to highest-scored chunks)
• Duplicate/overlapping chunks from the same source are merged
• Source attribution markers are embedded for citation extraction
• Unit tests for context assembly, token budgeting, and deduplication