fix: Remove OpenAI-specific message coalescing and consolidate token extraction

_coalesce_tool_messages_for_openai was a workaround for the parallel fan-out
problem that no longer exists with Command-based handoff routing. With
add_messages reducer and parallel_tool_calls=False, message state is always
well-ordered, so remove it along with its helper functions.

Consolidate token extraction in LDMetricsCallbackHandler to use
get_ai_usage_from_response from langchain_helper, reading from the
generation's AIMessage (usage_metadata / response_metadata) rather than
LLMResult.llm_output — the LangChain 1.x standard across all providers.

Fix TestBuildTools tests that were importing a nonexistent build_tools symbol;
update to build_structured_tools, covering both sync and async callables.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: jsonbailey

packages/ai-providers/server-ai-langchain/tests/test_tracking_langgraph.py

@@ -232,8 +232,11 @@ async def test_tracks_node_and_graph_tokens_on_success():
     handler.on_chain_start({}, {}, run_id=node_run_id, name='root-agent')
 
     llm_result = LLMResult(
-        generations=[[ChatGeneration(message=_AIMsg(content='Sunny.'), text='Sunny.')]],
-        llm_output={'token_usage': {'total_tokens': 15, 'prompt_tokens': 10, 'completion_tokens': 5}},
+        generations=[[ChatGeneration(
+            message=_AIMsg(content='Sunny.', usage_metadata={'total_tokens': 15, 'input_tokens': 10, 'output_tokens': 5}),
+            text='Sunny.',
+        )]],
+        llm_output={},
     )
     handler.on_llm_end(llm_result, run_id=uuid4(), parent_run_id=node_run_id)
     handler.on_chain_end({}, run_id=node_run_id)
@@ -389,8 +392,11 @@ async def test_tracks_graph_key_on_node_events():
     handler.on_chain_start({}, {}, run_id=node_run_id, name='root-agent')
 
     llm_result = LLMResult(
-        generations=[[ChatGeneration(message=_AIMsg(content='OK.'), text='OK.')]],
-        llm_output={'token_usage': {'total_tokens': 8, 'prompt_tokens': 5, 'completion_tokens': 3}},
+        generations=[[ChatGeneration(
+            message=_AIMsg(content='OK.', usage_metadata={'total_tokens': 8, 'input_tokens': 5, 'output_tokens': 3}),
+            text='OK.',
+        )]],
+        llm_output={},
     )
     handler.on_llm_end(llm_result, run_id=uuid4(), parent_run_id=node_run_id)
     handler.flush(graph, tracker)
@@ -459,16 +465,22 @@ def model_factory(node_config, **kwargs):
     root_run_id = uuid4()
     handler.on_chain_start({}, {}, run_id=root_run_id, name='root-agent')
     root_llm_result = LLMResult(
-        generations=[[ChatGeneration(message=_AIMsg(content='Root done.'), text='Root done.')]],
-        llm_output={'token_usage': {'total_tokens': 15, 'prompt_tokens': 10, 'completion_tokens': 5}},
+        generations=[[ChatGeneration(
+            message=_AIMsg(content='Root done.', usage_metadata={'total_tokens': 15, 'input_tokens': 10, 'output_tokens': 5}),
+            text='Root done.',
+        )]],
+        llm_output={},
     )
     handler.on_llm_end(root_llm_result, run_id=uuid4(), parent_run_id=root_run_id)
 
     child_run_id = uuid4()
     handler.on_chain_start({}, {}, run_id=child_run_id, name='child-agent')
     child_llm_result = LLMResult(
-        generations=[[ChatGeneration(message=_AIMsg(content='Child done.'), text='Child done.')]],
-        llm_output={'token_usage': {'total_tokens': 5, 'prompt_tokens': 3, 'completion_tokens': 2}},
+        generations=[[ChatGeneration(
+            message=_AIMsg(content='Child done.', usage_metadata={'total_tokens': 5, 'input_tokens': 3, 'output_tokens': 2}),
+            text='Child done.',
+        )]],
+        llm_output={},
     )
     handler.on_llm_end(child_llm_result, run_id=uuid4(), parent_run_id=child_run_id)