The protocol says nothing about the compositor implementation, because it is a protocol. X11 is the same way; ICCCM and EWMH are protocols that use X11 as an IPC mechanism. There is nothing stopping you from writing a multi-process compositor swarm with all the traditional pieces broken out using an IPC channel of your choice.

The fact is, given that a compositor must necessarily handle display and input, it is much more convenient (and secure) to centralize display and input dispatch in a single process. This is also how X11 works, except it dispatches all display and output to any process that asks nicely.