fix(mcp): preserve non-ASCII text in JSON serialization

json.dumps calls in MCP tool return_value_as_string and
StreamElicitor.elicit were missing ensure_ascii=False, causing
non-ASCII characters (Japanese, Chinese, Korean, etc.) to be
escaped as \uXXXX sequences. This degrades LLM performance
when working with non-English text since models handle native
characters better than unicode escapes.

Fixes #6995

Author: aj

python/packages/autogen-ext/src/autogen_ext/tools/mcp/_base.py

@@ -194,4 +194,4 @@ def serialize_item(item: Any) -> dict[str, Any]:
             else:
                 return {}
 
-        return json.dumps([serialize_item(item) for item in value])
+        return json.dumps([serialize_item(item) for item in value], ensure_ascii=False)

python/packages/autogen-ext/src/autogen_ext/tools/mcp/_host/_elicitation.py

@@ -82,7 +82,7 @@ async def elicit(self, params: mcp_types.ElicitRequestParams) -> mcp_types.Elici
                 prompt = "\n".join(
                     [
                         "Input Schema:",
-                        json.dumps(params.requestedSchema, indent=2),
+                        json.dumps(params.requestedSchema, indent=2, ensure_ascii=False),
                         "Please enter a JSON string following the above schema: ",
                     ]
                 )