Welcome to the fun of floating-point arithmetic. First, the duplicate points *are* removed, but this is not obvious unless you display all the digits of the result. For your third example, the four "duplicates" are actually all different, but only in the final bit (you can see it with "format hex"). I don't think there's anything I can do to make this work correctly in all cases.
Second, why should something fail if an intersection occurs at a segment boundary? Some people would like to consider this a proper intersection and others not (I know this from several discussions on the topic). As a user, you will have to evaluate the results you get for your particular application.
In many real world applications, where the two curves are formed from real data, this issue never comes up at all. It's (mostly) only from textbook type problems that you'll see this behavior.