Yup, I actually agree completely that having Julia as the server would be better. There are a couple other use cases like yours, but the main one is starting Julia remotely, and in particular you want Julia to keep running and be able to reconnect to it once your computation is done or whatever.
The main reason for the current model is that it just makes coordination simpler. Right now you start a server, boot Julia and it connects back. If Julia is the server then you have to boot Julia, have Julia tell Atom that it's listening, and then have Atom connect.
By the way, that's a nice patch, you should consider writing CoffeeScript / JS more often ;)