send_to_queue, wait_for_all_mots,
ClientCache) for managing concurrent model-output thunks; OpenAI-compatible
message conversion helpers (message_to_openai_message, messages_to_docs,
chat_completion_delta_merge); and _ServerType detection for adapting
structured-output support to the target server. Most user code will not import from
this package directly — it is consumed internally by the backend layer.