Implement hybrid retriever (vector + keyword merge)

[bug-ops]

Epic: #364

Description

Combine Qdrant vector search with SQLite FTS5 keyword search using weighted score merge.

Design

[memory.hybrid_search]
enabled = true
vector_weight = 0.7
keyword_weight = 0.3

Tasks

• HybridRetriever struct wrapping SemanticMemory + FTS5
• recall() runs both searches concurrently via tokio::try_join!
• Weighted merge: final_score = vector_weight * v_score + keyword_weight * k_score
• Deduplicate by message_id, keep highest score
• Config section [memory.hybrid_search] with weights
• Benchmark: hybrid vs vector-only recall quality