Add graph-aware ANN hybrid entity resolution

[JonasReuter]

Summary

Adds a new ann_hybrid entity resolution strategy that is designed to scale without curated alias/word arrays.

The strategy:

• builds graph-aware entity profile texts from title, type, description, neighbouring entities, and predicates
• embeds those profiles and uses the existing vector store for top-k ANN-style candidate blocking
• scores candidate pairs using embedding similarity, title similarity, graph-context overlap, and type compatibility
• auto-merges only high-confidence groups
• sends only borderline groups to the existing LLM merge extractor
• keeps the existing llm_context_window and embedding_search strategies intact

Config additions

entity_resolution:
  enabled: true
  strategy: ann_hybrid
  similarity_threshold: 0.72
  top_k: 10
  auto_merge_threshold: 0.95
  llm_review_threshold: 0.82
  max_llm_groups: 5000
  profile_neighbor_limit: 30

Notes

This first version avoids adding a new dependency such as rapidfuzz; it uses stdlib title similarity plus graph context and existing embeddings/vector store infrastructure.

Testing

Not run locally from this environment. Please run the repository's normal checks, especially:

uv run poe check
uv run poe test_unit