feat: Document Router Layer

[zTgx]

Problem

Workspace-scoped queries send all DocCards to the Orchestrator's LLM analysis phase in a single prompt. At 1000+ documents this exceeds token budgets, is expensive, and LLM selection quality degrades.

Proposal

Add a Router layer between Engine.query() and the Orchestrator that pre-filters documents using compile-stage artifacts (DocCard, ReasoningIndex, DocumentGraph) before LLM analysis.

• BM25 on DocCard text for lexical matching
• Keyword overlap using QueryPlan concepts vs ReasoningIndex topics
• Graph boost from cross-document relationships
• Optional LLM-assisted re-ranking of top candidates
• Only activates when workspace exceeds configurable threshold (default: 20 docs)

Full RFC

See docs/rfc/router

Acceptance Criteria

• rust/src/router/ module with DocumentRouter, RouteResult, RouterConfig
• Three-signal scoring (BM25 + keyword + graph)
• Incremental index maintenance on document add/remove
• Integration into Engine.query() with activation threshold
• RouterConfig exposed to Python SDK
• Unit tests + 1000-document simulation test

Open Questions

• Score normalization strategy across different corpus sizes
• Cold start handling for new documents (no graph edges)
• Thread safety model for mutable BM25 index (RwLock vs rebuild-on-query)