fix(workflow): track fixed speech at transport output

backend/services/workflow/realtime.py

@@ -23,6 +23,7 @@ from services.message_stage import (
     MessageStageRunner,
     MessageStageSpec,
 )
+from services.pipecat.call_lifecycle import playback_marker_for
 from services.pipecat.realtime_tools import RealtimeTool, RealtimeToolResult
 from services.runtime_variables import DynamicVariableError, DynamicVariableStore
 from services.system_tools import state_update_properties, system_tool_kind
@@ -99,6 +100,18 @@ class RealtimeWorkflowOutput(WorkflowOutput):
             content,
             suppress_transcript=True,
         )
+        playback_marker = playback_marker_for(playback_completion)
+        if playback_marker is not None:
+            async def finish_playback_tracking() -> None:
+                # The provider boundary is emitted after its final audio frame.
+                # Queueing the marker then places it behind that audio at output.
+                if provider_completion is not None:
+                    await provider_completion
+                await self._runtime.queue_frame(playback_marker)
+                playback_marker.completion.mark_queued()
+                await playback_completion
+
+            return asyncio.create_task(finish_playback_tracking())
         # Message playback policy and deterministic continuation must use the
         # transport boundary.  Provider response.done only means generation
         # has finished; audio may still be buffered at the output transport.
@@ -983,12 +996,20 @@ class WorkflowRealtimeController:
             return
         self._runtime.call_end.begin("workflow_completed")
         if message:
-            self._runtime.call_end.arm_after_speech()
             completion = await self._output.speak(
                 message,
                 source="workflow-end-speech",
                 node_id=node_id,
             )
+            arm_tracked = getattr(
+                self._runtime.call_end,
+                "arm_after_tracked_speech",
+                None,
+            )
+            if callable(arm_tracked):
+                await arm_tracked()
+            else:
+                self._runtime.call_end.arm_after_speech()
             if completion:
                 await completion
         else: