Cross repo isolation

tests/test_llm_extraction.py

@@ -50,6 +50,58 @@ def chunk():
     )
 
 
+def test_llm_prompt_filters_derived_scope_and_labels_source_roles():
+    adapter = FakeAdapter('{"abilities": []}')
+    extractor = LLMCandidateExtractor(adapter)
+    chunks = [
+        ContentChunk(
+            id=1,
+            repository_id=1,
+            analysis_run_id=1,
+            snapshot_id=1,
+            path="SCOPE.md",
+            kind="scope",
+            start_line=1,
+            end_line=3,
+            text="# SCOPE\n\nOld approved LLM routing entry.",
+            metadata={"source_role": "derived_scope"},
+        ),
+        ContentChunk(
+            id=2,
+            repository_id=1,
+            analysis_run_id=1,
+            snapshot_id=1,
+            path="INTENT.md",
+            kind="intent",
+            start_line=1,
+            end_line=3,
+            text="# INTENT\n\nProvide lightweight IAM.",
+            metadata={"source_role": "intent_summary"},
+        ),
+        ContentChunk(
+            id=3,
+            repository_id=1,
+            analysis_run_id=1,
+            snapshot_id=1,
+            path="CLAUDE.md",
+            kind="documentation",
+            start_line=1,
+            end_line=2,
+            text="# CLAUDE\n\nAgent guidance.",
+            metadata={"source_role": "agent_guidance"},
+        ),
+    ]
+
+    extractor.extract(repository(), chunks)
+
+    assert "Source: INTENT.md" in adapter.last_prompt
+    assert "source_role=intent_summary" in adapter.last_prompt
+    assert "Source: SCOPE.md" not in adapter.last_prompt
+    assert "Old approved LLM routing entry" not in adapter.last_prompt
+    assert "Source: CLAUDE.md" not in adapter.last_prompt
+    assert "Do not use SCOPE.md" in adapter.last_prompt
+
+
 def test_llm_candidate_extractor_parses_structured_response():
     adapter = FakeAdapter(
         """