Summary

Phase 3 of graph memory (#1222): augment existing recall with graph traversal, scoring, and RRF fusion.

Depends on: #1225 (extraction pipeline)

Tasks

1. Graph Retrieval (graph/retrieval.rs)

graph_recall(query: &str, store: &GraphStore, qdrant: &EmbeddingStore, provider: &AnyProvider, config: &GraphConfig) -> Result<Vec<GraphFact>>:

1. Entity matching: Embed query → search zeph_graph_entities Qdrant collection (top-5 entities by cosine). Also run SQLite fuzzy match as fallback.
2. BFS expansion: For each matched entity, run store.bfs(entity_id, config.max_hops) to collect neighbors + edges.
3. Scoring: For each collected edge, compute GraphFact.composite_score() = entity_match_score * (1/(1+hop_distance)) * confidence.
4. Dedup: Remove duplicate edges (same source+target+relation).
5. Rank and truncate: Sort by composite_score desc, take top config.recall_limit.

2. Router Update

Add Graph variant to MemoryRoute enum (always present, not cfg-gated). Router only returns Graph when graph-memory feature is enabled. When disabled, Graph route is treated as Hybrid.

Update HeuristicRouter:

• "What does the user think about X?" → Graph route
• "How is X related to Y?" → Graph route
• Entity-centric queries (contains known entity name) → Graph route

3. Context Budget

Add graph_facts field to BudgetAllocation (always present, 0 when disabled):

• Redistribute: code_context 30% → 29%, graph_facts 0% → 4% (only when enabled)
• Existing allocations: summaries=8%, semantic_recall=8%, cross_session=4%, recent_history=50%

4. Context Formatting

Format graph facts as a system message with GRAPH_FACTS_PREFIX:

## Known Facts
- User prefers neovim for Rust development (since 2026-01) [confidence: 0.95]
- User works on project "zeph" (AI agent) [confidence: 1.0]

Architecture Reference

See .local/plan/graph-memory-architecture.md Section 5 for retrieval pseudocode, router patterns, and budget allocation.

Acceptance Criteria

• Graph recall returns relevant facts for entity-centric queries
• BFS traversal via recursive CTE respects max_hops
• RRF fusion produces better results than vector-only on relationship queries
• Router correctly identifies graph-suitable queries
• Budget allocation correct when enabled/disabled
• Context formatted with entity names, relations, and confidence
• ~20 tests (18 unit + 2 integration)