I agree that the proposed behaviour seems more correct:  the collection of 
all subsets of size 4 of range(3) is perfectly valid and well-defined;  it 
just happens to be empty.  I've also encountered this in practice (in code 
that was enumerating partitions of a list into a fixed number of pieces), 
but in my case it was easier to work around.

I guess the counterargument is that the current ValueError might catch 
bugs early;  I don't know how often this is true in practice.

In any case, this is Raymond's call.