feat: structure-aware retrieval scoring using knowledge graph

[salishforge]

Summary

Incorporate knowledge graph relationships into retrieval ranking. A memory that connects to entities mentioned in the query should rank higher than one with equal text similarity but no graph connections.

Design

After standard retrieval (FTS/semantic/hybrid), apply a graph-aware boost:

• Extract entity names from the query (regex: capitalized multi-word, known entity names)
• For each result, check how many of its linked entities overlap with query entities
• Apply multiplicative boost: score *= (1 + 0.2 * entity_overlap_fraction)
• Zero LLM cost — uses existing warm_tier_entities + entities tables

Inspiration

Adapted from FABLE (arXiv 2601.18116) structure-aware propagation S(v) = 1/3[S_sim + S_inh + S_child]. Simplified to use MemForge's existing knowledge graph instead of document trees.