[ARCH] Replace flat registry with true graph data model (nodes + edges)

[Wolfvin]

Problem

CodeLens currently uses a flat registry (SQLite rows per symbol). This works for simple lookup but blocks structural queries like: who calls this function across the entire codebase, what is the blast radius if I rename this class, are there circular dependency chains.

Many existing engines (trace, impact, circular) reimplement partial graph traversal ad-hoc, leading to duplicated logic.

Proposed Change

Adopt a proper node + edge graph model backed by SQLite:

• Nodes: Function, Class, File, Module, Route, Type, Interface
• Edges: CALLS, IMPORTS, DEFINES, INHERITS, IMPLEMENTS, USES_TYPE
• Each edge stores: source_id, target_id, edge_type, file, line, confidence

Benefits for Agents

• Single traversal primitive replaces all ad-hoc engine code
• Accurate blast radius: agents know exactly what breaks before touching anything
• Foundation for Cypher-like queries (see separate issue)

Migration Path

1. Add nodes + edges tables alongside existing registry (non-breaking)
2. Populate during scan pass
3. Migrate engines one by one
4. Deprecate flat registry after all engines migrated

Reference

codebase-memory-mcp uses this model and achieves <1ms query time on 4.8M nodes via SQLite with proper indexing.