Monitor and server tasks are never cancelled on shutdown, causing non-zero exit after training completes

[Fiona-Waters]

Description

Two fire-and-forget asyncio tasks are never cancelled during backend shutdown:

1. _monitor_openai_server in LocalBackend (src/art/local/backend.py, line ~491) — created via asyncio.create_task() but the reference is never stored. The monitor polls vLLM health every 30s. After vLLM shuts down, it hits 3 consecutive ConnectionRefusedError failures and re-raises, producing a "Task exception was never retrieved" error and a non-zero process exit.
2. openai_server_task in UnslothService (src/art/unsloth/service.py, line ~516) — in shared mode, the returned asyncio task (the uvicorn server) is discarded. It continues running as an orphaned task after shutdown.

Impact

In containerized environments (Kubeflow TrainJob, KubeRay), the non-zero exit causes the pod to fail/restart even though training completed successfully — adapter checkpoints, metrics, and results are all persisted to disk before the error occurs.

Steps to reproduce

1. Run a LocalBackend training loop in shared mode (in_process=True) inside a container
2. Let training complete normally
3. Observe ConnectionRefusedError in the asyncio task exception output
4. Process exits non-zero

Error output

  Traceback (most recent call last):
    File "/usr/lib64/python3.12/asyncio/tasks.py", line 314, in __step_run_and_handle_result
      result = coro.send(None)
               ^^^^^^^^^^^^^^^
    File "/opt/app-root/lib64/python3.12/site-packages/art/local/backend.py", line 436, in _monitor_openai_server
      async with session.get(
                 ^^^^^^^^^^^^
    File "/opt/app-root/lib64/python3.12/site-packages/aiohttp/client.py", line 1521, in __aenter__
      self._resp: _RetType = await self._coro
                             ^^^^^^^^^^^^^^^^
    File "/opt/app-root/lib64/python3.12/site-packages/aiohttp/client.py", line 788, in _request
      resp = await handler(req)
             ^^^^^^^^^^^^^^^^^^
    File "/opt/app-root/lib64/python3.12/site-packages/aiohttp/client.py", line 742, in _connect_and_send_request
      conn = await self._connector.connect(
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    File "/opt/app-root/lib64/python3.12/site-packages/aiohttp/connector.py", line 672, in connect
      proto = await self._create_connection(req, traces, timeout)
              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    File "/opt/app-root/lib64/python3.12/site-packages/aiohttp/connector.py", line 1251, in _create_connection
      _, proto = await self._create_direct_connection(req, traces, timeout)
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    File "/opt/app-root/lib64/python3.12/site-packages/aiohttp/connector.py", line 1623, in _create_direct_connection
      raise last_exc
    File "/opt/app-root/lib64/python3.12/site-packages/aiohttp/connector.py", line 1592, in _create_direct_connection
      transp, proto = await self._wrap_create_connection(
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    File "/opt/app-root/lib64/python3.12/site-packages/aiohttp/connector.py", line 1333, in _wrap_create_connection
      raise client_error(req.connection_key, exc) from exc
  aiohttp.client_exceptions.ClientConnectorError: Cannot connect to host 0.0.0.0:54993 ssl:default [Connect call failed ('0.0.0.0', 54993)]

Environment

• openpipe-art 0.5.17
• vLLM 0.18.x (V1 engine only)