The point is that you can't make a second overlapping noise pattern without changing the first. You can't yell anything that turns an existing yell into silence.
You're bitwise-anding all the yelling together, and only if that validates the key you got, do you trust it.
So MITM can't yell while the router's yelling unless its hash starts the way the router's ends (and it knows this in time and starts yelling at the right time) or the client will see a bad hash, tear down and try again.
And if the MITM yells _after_ the router, then it's too late, the client has already gotten a key with a valid hash yelled with it, and is secure.
And if the MITM tries to drown out the yelling (or the key) from the router, the client can see this unusually long yell, and know that a key was being sent at the time, and will tear down and try again.
This was on page 2 of the article...