Delegated conversations clean up reliably

When one agent asks another agent for help, the mesh sets up a small watch that waits for that delegated sub-conversation to finish so it can release its resources. In one narrow timing case, the watch could fail to recognise that the sub-conversation was actually done, leaving it running indefinitely instead of releasing it. Left unnoticed for long enough this could accumulate and slow the whole system down.

The watch now recognises a finished sub-conversation correctly every time, so delegated conversations always release their resources once they are done.