fix(langchain): detach existing SpanHolder token before overwrite in …

…_create_llm_span

Closes #3957

When _create_llm_span() is called for a run_id that already has an entry
in self.spans, the old holder's token was lost without being detached,
leaving an orphaned context_api.attach() on the OTel context stack. This
is the same class of bug as #3526 / #3807.

Defensively detach the existing holder's token before replacing the entry.

packages/opentelemetry-instrumentation-langchain/opentelemetry/instrumentation/langchain/callback_handler.py

@@ -472,6 +472,13 @@ def _create_llm_span(
             # If context setting fails, continue without suppression token
             token = None
 
+        # Detach any existing holder's token before overwriting, otherwise
+        # the previous context_api.attach() is orphaned and corrupts the
+        # OTel context stack (same class of bug as #3526 / #3807).
+        existing_holder = self.spans.get(run_id)
+        if existing_holder is not None and existing_holder.token is not None:
+            self._safe_detach_context(existing_holder.token)
+
         self.spans[run_id] = SpanHolder(
             span, token, None, [], workflow_name, None, entity_path
         )