fix: reset empty response retry counter after successful recovery

The retry counter was per-invocation, not per-failure-burst. If a model
returned empty responses at different points during the same invocation,
earlier (recovered) empties consumed the budget. A later empty response
would exhaust the counter and halt silently.

Fix: reset empty_response_count to 0 after any successful (non-empty)
response. Also add a warning log when retries are exhausted so the halt
is not silent.

Author: gurjot-05

src/google/adk/flows/llm_flows/base_llm_flow.py

@@ -808,6 +808,11 @@ async def run_async(
       ):
         is_empty_response = True
 
+      # Reset retry budget after a successful (non-empty) response so that
+      # later empty responses in the same invocation still get retried.
+      if not is_empty_response and empty_response_count > 0:
+        empty_response_count = 0
+
       if (
           is_empty_response
           and empty_response_count < _MAX_EMPTY_RESPONSE_RETRIES
@@ -841,6 +846,16 @@ async def run_async(
         )
         continue
 
+      if (
+          is_empty_response
+          and empty_response_count >= _MAX_EMPTY_RESPONSE_RETRIES
+      ):
+        logger.warning(
+            'Model returned an empty response but retry budget (%d) is'
+            ' exhausted. Halting.',
+            _MAX_EMPTY_RESPONSE_RETRIES,
+        )
+
       # Normal termination conditions.
       if not last_event or last_event.is_final_response() or last_event.partial:
         if last_event and last_event.partial: