(this post is in the context of Ethernet, other procols may vary)
What criteria is used to know the boundaries of a collision domain?
A collision domain is an area of a network where collisions can happen and where only one packet may be successfully transmitted at a time. It passes through repeaters (hubs) but not through bridges (switches).
All half duplex ethernet connections are part of a collision domain, that domain may be limited to a single physical layer connection (which may be either point to point or multipoint) or it may pass through repeaters.
Full duplex connections must be point to point and can't connect to a repeater. In this case there is simply no possibility of collison and hence no collision domain.
Why does an ethernet switch break up a collision domain?
Repeaters (hubs) work at a low level, they do not have the ability to buffer frames or to identify the intended destinition of a frame. So if they detect a collision then all they can do is output jam signals to ensure every device in the collision domain sees the collision.
Bridges (switches) work at a higher level. They process complete frames and forward them based on destination mac address. If a bridge detects a collision it will perform the CSMA/CD procedure just like an end device would. There is no need to repeat the collision out of other ports.
Are broadcast domains and collision domains mutually exclusive?
They are different concepts at different layers. A broadcast domain may cover multiple collision domains, or (if all the links are full duplex) there may be no collision domains at all but there is still a broadcast domain.