Summary

Follow-up to graph memory epic (#1222). MVP uses exact name+type match for entity resolution. This issue adds embedding-based resolution for higher quality deduplication.

When use_embedding_resolution = true:

• Embed entity name+summary → search existing entities via Qdrant
• cosine > 0.85: merge entities
• 0.70-0.85: LLM disambiguation
• < 0.70: create new entity

Batch embedding to respect timeout budget.

Depends on: #1225 (extraction pipeline)