fix: Resolve HIGH priority grounding_metadata state management issues

- Add interrupted parameter to __build_full_text_response to preserve interrupted signal when flushing pending text
- Pass interrupted flag in turn_complete and interrupted blocks
- Remove premature reset of last_grounding_metadata after interrupted (not a terminal event)
- Add documentation for tool_call metadata persistence design decision

Addresses review comments:
- HIGH: Lost interrupted signal in full text response
- HIGH: Premature reset after interrupted
- MEDIUM: Duplicate reset logic (simplified by removing premature reset)

Author: VedantMadane

src/google/adk/models/gemini_llm_connection.py

@@ -143,6 +143,7 @@ def __build_full_text_response(
       self,
       text: str,
       grounding_metadata: types.GroundingMetadata | None = None,
+      interrupted: bool = False,
   ):
     """Builds a full text response.
 
@@ -152,6 +153,7 @@ def __build_full_text_response(
     Args:
       text: The text to be included in the response.
       grounding_metadata: Optional grounding metadata to include.
+      interrupted: Whether this response was interrupted.
 
     Returns:
       An LlmResponse containing the full text.
@@ -162,6 +164,7 @@ def __build_full_text_response(
             parts=[types.Part.from_text(text=text)],
         ),
         grounding_metadata=grounding_metadata,
+        interrupted=interrupted,
     )
 
   async def receive(self) -> AsyncGenerator[LlmResponse, None]:
@@ -295,7 +298,9 @@ async def receive(self) -> AsyncGenerator[LlmResponse, None]:
           if message.server_content.turn_complete:
             if text:
               yield self.__build_full_text_response(
-                  text, last_grounding_metadata
+                  text,
+                  last_grounding_metadata,
+                  interrupted=message.server_content.interrupted,
               )
               text = ''
             yield LlmResponse(
@@ -312,15 +317,14 @@ async def receive(self) -> AsyncGenerator[LlmResponse, None]:
           if message.server_content.interrupted:
             if text:
               yield self.__build_full_text_response(
-                  text, last_grounding_metadata
+                  text, last_grounding_metadata, interrupted=True
               )
               text = ''
             else:
               yield LlmResponse(
                   interrupted=message.server_content.interrupted,
                   grounding_metadata=last_grounding_metadata,
               )
-            last_grounding_metadata = None  # Reset after yielding
         if message.tool_call:
           if text:
             yield self.__build_full_text_response(text, last_grounding_metadata)
@@ -333,6 +337,10 @@ async def receive(self) -> AsyncGenerator[LlmResponse, None]:
               content=types.Content(role='model', parts=parts),
               grounding_metadata=last_grounding_metadata,
           )
+          # Note: last_grounding_metadata is NOT reset here because tool_call
+          # is part of an ongoing turn. The metadata persists until turn_complete
+          # or interrupted with break, ensuring subsequent messages in the same
+          # turn can access the grounding information.
         if message.session_resumption_update:
           logger.debug('Received session resumption message: %s', message)
           yield (