fix(core): clear usage metadata callback on exceptions in context manager (#39616)

Co-authored-by: Solaris-star <820622658@qq.com>

libs/core/langchain_core/callbacks/usage.py

@@ -114,6 +114,11 @@ def get_usage_metadata_callback(
     )
     register_configure_hook(usage_metadata_callback_var, inheritable=True)
     cb = UsageMetadataCallbackHandler()
-    usage_metadata_callback_var.set(cb)
-    yield cb
-    usage_metadata_callback_var.set(None)
+    token = usage_metadata_callback_var.set(cb)
+    try:
+        yield cb
+    finally:
+        # Always clear the context var, including when the with-block raises.
+        # Without finally, post-block model calls keep accumulating into cb
+        # (see #38989).
+        usage_metadata_callback_var.reset(token)

libs/core/tests/unit_tests/callbacks/test_usage_callback.py

@@ -1,3 +1,4 @@
+import contextlib
 from typing import Any
 
 from langchain_core.callbacks import (
@@ -120,3 +121,15 @@ async def test_usage_callback_async() -> None:
     callback = UsageMetadataCallbackHandler()
     _ = await llm.abatch(["Message 1", "Message 2"], config={"callbacks": [callback]})
     assert callback.usage_metadata == {"test_model": total_1_2}
+
+
+def test_usage_callback_clears_on_exception() -> None:
+    """Callback must stop tracking after with-block exits via exception (#38989)."""
+    llm = FakeChatModelWithResponseMetadata(messages=iter(messages), model_name="fake")
+    with contextlib.suppress(RuntimeError), get_usage_metadata_callback() as cb:
+        _ = llm.invoke("in block")
+        raise RuntimeError
+
+    # Calls after the block must not accumulate into the previous callback.
+    _ = llm.invoke("outside block")
+    assert cb.usage_metadata == {"fake": usage1}