Demonstrates append_to_context=True for intro lines and the
on_function_calls_started + pause_frame_processing=True pattern
for tool-call filler, so injected speech lands in the transcript
in the correct turn order without overlapping the post-tool
LLM response.
2243cc11b9