feat(memory): MAGMA multi-graph memory with typed edges (issue #1821)

[bug-ops]

Summary

Implement MAGMA (Multi-Graph Agentic Memory Architecture) — multi-graph memory with typed edges (Semantic/Temporal/Causal/Entity) enabling subgraph-scoped BFS retrieval.

4-Phase Implementation:

1. Schema: EdgeType enum + migration 041 with UNIQUE index fix + CHECK constraint
2. Extraction: LLM edge_type classification with serde defaults + fallback to Semantic
3. Retrieval: Query classifier (heuristic) + typed subgraph BFS with dynamic filtering
4. Integration: Context assembly calls classifier, TUI/CLI display edge types

Test Plan

• 6281/6281 tests PASS (+17 new tests)
• All validators approved (Security, Impl-Critic, Perf, Tester)
• Code review approved with 5 minors (all fixed)
• fmt, clippy, nextest all PASS
• Both critic mitigations verified and tested
• Backward compatibility confirmed (empty edge_types = all types)

Key Changes

• Migration 041: edge_type column + FIXED UNIQUE index + perf indexes + CHECK constraint
• EdgeType enum: Semantic(default), Temporal, Causal, Entity with serde, Display, FromStr
• insert_edge_typed(): dedup key includes edge_type (critic mitigation M0: Project bootstrap — workspace and crate skeleton #1)
• bfs_typed(): typed subgraph BFS with dynamic IN filtering
• graph_recall(): edge_types parameter, dedup key includes edge_type (critic mitigation M1: Ollama chat loop — interactive CLI with LLM #2)
• classify_graph_subgraph(): heuristic query classifier with shared marker constants
• Context assembly: calls classify_graph_subgraph per query
• LLM extraction: ExtractedEdge.edge_type with serde defaults + parse fallback
• TUI/CLI: /graph stats shows edge_type_distribution, /graph facts/history display types

Validation Results

• Security: 0 vulnerabilities, LLM injection defense solid, SQL injection prevention correct
• Implementation Logic: All 8 focus areas correct (dedup, BFS, classifier, migration, resolver, error handling, backward compatibility)
• Performance: No regressions, classify <1µs, BFS indexed, memory overhead negligible
• Test Coverage: 6281/6281 PASS, all mitigations verified, edge cases covered
• Code Quality: fmt/clippy/nextest all PASS, CHANGELOG updated, architecture adherent

Related Issues

Closes #1821
Related: #1817 (A-MEM note linking), #1341 (temporal edge columns), #1789 (TA-Mem adaptive retrieval)