fix(agents): cite the pages the graph tool read

The graph tool records every page read_entity_pages returns in
retrieved_docs, but agents only ever collected internal_search's. A turn
answered from graph pages therefore emitted no sources: on the multi-hop e2e
run, a question answered from two read_entity_pages calls came back with none.

_search_tool_docs gathers both tools' documents the way the executor caches
them, and both the classic/agentic collector and the research agent's
per-step citations use it. Checked against the real executor and a real
graph: a graph-only turn now cites the pages it read.

tests/agents/test_classic_agent.py

@@ -313,6 +313,26 @@ class TestClassicAgentSearchExposure:
             "Tool Doc",
         ]
 
+    def test_collect_internal_sources_includes_graph_pages(
+        self, agent_base_params, mock_llm_creator, mock_llm_handler_creator
+    ):
+        # Pages the graph tool read carry the answer as much as search hits do,
+        # so they are cited the same way.
+        from docsgpt.agents.tools.graph_search import GRAPH_TOOL_ID
+
+        retriever_config = {"source": {"active_docs": ["b"]}}
+        agent = ClassicAgent(retriever_config=retriever_config, **agent_base_params)
+        search = Mock()
+        search.retrieved_docs = [{"text": "Found", "title": "Search Doc", "source": "b"}]
+        graph = Mock()
+        graph.retrieved_docs = [{"text": "Quill is a store.", "title": "quill.md", "source": "b"}]
+        user = agent.user or ""
+        agent.tool_executor._loaded_tools[f"internal_search:{INTERNAL_TOOL_ID}:{user}"] = search
+        agent.tool_executor._loaded_tools[f"graph_search:{GRAPH_TOOL_ID}:{user}"] = graph
+
+        agent._collect_internal_sources()
+        assert [d["title"] for d in agent.retrieved_docs] == ["Search Doc", "quill.md"]
+
     def test_collect_internal_sources_dedupes(
         self, agent_base_params, mock_llm_creator, mock_llm_handler_creator
     ):

tests/agents/test_research_agent.py

@@ -779,6 +779,20 @@ class TestCollectStepSources:
 
         assert len(agent.citations.citations) == 2
 
+    def test_collects_pages_the_graph_tool_read(
+        self, agent_base_params, mock_llm_creator, mock_llm_handler_creator
+    ):
+        from docsgpt.agents.tools.graph_search import GRAPH_TOOL_ID
+
+        agent = ResearchAgent(**agent_base_params)
+        graph = Mock()
+        graph.retrieved_docs = [{"source": "s3", "title": "quill.md", "text": "Quill"}]
+        agent.tool_executor._loaded_tools[f"graph_search:{GRAPH_TOOL_ID}:{agent.user or ''}"] = graph
+
+        agent._collect_step_sources()
+
+        assert len(agent.citations.citations) == 1
+
     def test_no_tool_no_error(
         self, agent_base_params, mock_llm_creator, mock_llm_handler_creator
     ):