fix: enforce consistent edge behavior across all backends

NetworkX auto-creates phantom nodes when adding edges between
non-existent nodes. SQLite stored dangling edges. Both now skip
edges where source or target nodes don't exist, matching KuzuDB's
FOREIGN KEY behavior.

All 3 backends produce identical results: same nodes, same edges.
Verified on contoso-real-estate: 2,298 nodes, 2,890 edges across
NetworkX, SQLite, and KuzuDB.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: aksOps

src/code_intelligence/graph/backends/networkx.py

@@ -38,6 +38,14 @@ def add_node(self, node: GraphNode) -> None:
         self._g.add_node(node.id, **node.model_dump())
 
     def add_edge(self, edge: GraphEdge) -> None:
+        # Only add edge if both nodes exist — prevents NetworkX from
+        # auto-creating phantom nodes for dangling references.
+        if edge.source not in self._g or edge.target not in self._g:
+            logger.debug(
+                "Skipping edge %s -> %s: missing node(s)",
+                edge.source, edge.target,
+            )
+            return
         self._g.add_edge(edge.source, edge.target, **edge.model_dump())
 
     def clear(self) -> None:

src/code_intelligence/graph/backends/sqlite_backend.py

@@ -127,6 +127,9 @@ def add_node(self, node: GraphNode) -> None:
             logger.exception("Failed to add node %s", node.id)
 
     def add_edge(self, edge: GraphEdge) -> None:
+        # Only add edge if both nodes exist — consistent with KuzuDB/Neo4j behavior
+        if not self.has_node(edge.source) or not self.has_node(edge.target):
+            return
         try:
             self._conn.execute(
                 "INSERT INTO edges (source, target, kind, data) VALUES (?, ?, ?, ?)",