docs: add semantic search roadmap

Technical roadmap for vector search, hybrid ranking, and RAG:

Architecture:
- VectorStore provider pattern (custom + LanceDB baseline)
- Lean interface returning IDs, orchestrator hydrates
- Three chunking strategies (line, sentence, semantic)
- Embedding provider abstraction (Ollama, OpenAI, Gemini, Transformers.js)
- LLM provider abstraction for RAG/reranking
- Hybrid search with FTS5 + RRF fusion
- Reranker preserves retrieval scores
- Evaluation with per-category breakdown

Phases:
- Phase 1: Vector search core (chunking, embeddings, brute-force)
- Phase 2: Hybrid search (FTS5, BM25, RRF, HNSW)
- Phase 3: RAG & evaluation (reranking, MCP integration, metrics)

Includes complete type definitions and CLI reference.

Author: prosdev