I think a better argument for twos complement is that you're just doing all of your computations modulo 2^n (where n is 32 or 64 or whatever), and addition and multiplication work as expected modulo anything.  The idea of negative and positive in a system like this is artificial, though.  The standard convention follows from the observation that 0-1 should be -1, and if you use the hardware that is for positive numbers, and assume that you can always carry another 1 for the highest order bit, then 0-1 is 1111...111.  It's not that 2^n - 1 is actually -1, but that it acts like a -1 would act.  It's only by convention that numbers which begin with a 1 are considered to be negative (that, and Intel has special hardware for detecting if a number has its leading 1 set).

However, we could adopt the convention that you need two leading ones to make a negative number, and everything would work out, except for the inequality testing built into the CPU.  This would give us a lot more positive numbers, and then abs wouldn't have this problem :-)

Or, three other options: 1) make MIN_INT outside the domain of abs, 2) make the range of abs be some unsigned int type, or 3) use Integer (i.e., use a type which actually represents integers rather than a type which can only handle small integers).

Kyle