diff --git a/graphify/diagnostics.py b/graphify/diagnostics.py
index fcb9a11cf..32896494b 100644
--- a/graphify/diagnostics.py
+++ b/graphify/diagnostics.py
@@ -161,7 +161,12 @@ def diagnose_extraction(
     max_examples: int = 5,
     extract_path: str | Path | None = None,
 ) -> dict[str, Any]:
-    """Summarize same-endpoint edge-collapse risk for one JSON graph/extraction dict."""
+    """Summarize graph-integrity risks for one JSON graph/extraction dict.
+
+    Missing ``ref_*`` targets are expected external references produced for
+    dependencies outside the scanned corpus. They are counted separately from
+    actionable internal dangling endpoints.
+    """
     from graphify.build import build_from_json
 
     node_ids = _node_ids(extraction)
@@ -184,6 +189,7 @@ def diagnose_extraction(
     non_object_edges = 0
     missing_endpoint_edges = 0
     dangling_endpoint_edges = 0
+    external_reference_edges = 0
     self_loop_edges = 0
     valid_candidate_edges = 0
 
@@ -197,7 +203,10 @@ def diagnose_extraction(
             missing_endpoint_edges += 1
             continue
         if source not in node_ids or target not in node_ids:
-            dangling_endpoint_edges += 1
+            if source in node_ids and target.startswith("ref_"):
+                external_reference_edges += 1
+            else:
+                dangling_endpoint_edges += 1
             continue
         if source == target:
             self_loop_edges += 1
@@ -252,6 +261,7 @@ def diagnose_extraction(
         "non_object_edges": non_object_edges,
         "missing_endpoint_edges": missing_endpoint_edges,
         "dangling_endpoint_edges": dangling_endpoint_edges,
+        "external_reference_edges": external_reference_edges,
         "self_loop_edges": self_loop_edges,
         "valid_candidate_edges": valid_candidate_edges,
         "exact_duplicate_edges": _count_extra(exact_counts),
@@ -358,6 +368,7 @@ def format_diagnostic_report(summary: dict[str, Any]) -> str:
         f"valid_candidate_edges: {summary['valid_candidate_edges']}",
         f"missing_endpoint_edges: {summary['missing_endpoint_edges']}",
         f"dangling_endpoint_edges: {summary['dangling_endpoint_edges']}",
+        f"external_reference_edges: {summary.get('external_reference_edges', 0)}",
         f"self_loop_edges: {summary['self_loop_edges']}",
         f"exact_duplicate_edges: {summary['exact_duplicate_edges']}",
         f"directed_unique_endpoint_pairs: {summary['directed_unique_endpoint_pairs']}",
diff --git a/tests/test_multigraph_diagnostics.py b/tests/test_multigraph_diagnostics.py
index 8c39b8e23..0206a0c13 100644
--- a/tests/test_multigraph_diagnostics.py
+++ b/tests/test_multigraph_diagnostics.py
@@ -103,6 +103,37 @@ def test_diagnose_extraction_categorizes_same_endpoint_collapse() -> None:
     assert summary["post_build_edge_count"] == 2
 
 
+def test_diagnose_extraction_separates_external_reference_edges() -> None:
+    extraction = {
+        "nodes": [
+            {"id": "app", "label": "app", "file_type": "code", "source_file": "app.ts"},
+        ],
+        "edges": [
+            {
+                "source": "app",
+                "target": "ref_react",
+                "relation": "imports_from",
+                "confidence": "EXTRACTED",
+                "source_file": "app.ts",
+            },
+            {
+                "source": "app",
+                "target": "missing_internal",
+                "relation": "references",
+                "confidence": "INFERRED",
+                "source_file": "app.ts",
+            },
+        ],
+    }
+
+    summary = diagnose_extraction(extraction, directed=True)
+    report = format_diagnostic_report(summary)
+
+    assert summary["external_reference_edges"] == 1
+    assert summary["dangling_endpoint_edges"] == 1
+    assert "external_reference_edges: 1" in report
+
+
 def test_diagnose_extraction_accepts_node_link_links_key() -> None:
     extraction = _diagnostic_fixture()
     extraction["links"] = extraction.pop("edges")