fix(cli): prevent 'Press ENTER to continue...' on exit

[teknium1]

Problem

During normal CLI usage (not on exit), users see this interrupting their session:

Unhandled exception in event loop:
  File httpx/_client.py in aclose
  ...
  RuntimeError: Event loop is closed

Press ENTER to continue...

Root Cause

The OpenAI SDK wraps httpx's AsyncClient in AsyncHttpxClientWrapper which has a __del__:

class AsyncHttpxClientWrapper(DefaultAsyncHttpxClient):
    def __del__(self) -> None:
        if self.is_closed:
            return
        try:
            asyncio.get_running_loop().create_task(self.aclose())
        except Exception:
            pass

The mid-session flow that triggers this:

1. Turn 1: Agent tool call runs on worker thread → creates AsyncOpenAI client bound to thread's event loop → client is cached in _client_cache
2. Worker thread finishes → its event loop closes
3. Turn 2: New worker thread → _get_cached_client() detects stale entry (loop is closed) → evicts it with del
4. del triggers __del__ → asyncio.get_running_loop() finds prompt_toolkit's running loop → schedules aclose()
5. aclose() tries to close TCP connections on the dead worker loop → RuntimeError: Event loop is closed
6. prompt_toolkit catches the unhandled exception and shows "Press ENTER to continue..."

Fix

• _force_close_async_httpx() — sets httpx AsyncClient._state = ClientState.CLOSED, so __del__ checks is_closed and returns immediately (no-op)
• Stale client eviction (_get_cached_client) — now calls _force_close_async_httpx() before del, preventing the mid-session error
• shutdown_cached_clients() — called from _run_cleanup() on CLI exit to clean up any remaining cached clients
• CLI cleanup — _run_cleanup() in cli.py calls shutdown_cached_clients()

Files Changed

• agent/auxiliary_client.py — new shutdown functions + fix stale eviction
• cli.py — call shutdown in cleanup