This isn't the case in this guys problem. He's just trying to figure out why the count he measures from the encoder don't match up with the pulses he sends to the stepper...
Well it could very well be. He states somewhere that he is simply 'counting pulses on 'a'. Thus every double reversal type 1 and 3 (a change, b steady) will give him an extra count.

You really lost me on this comment...what's a "double reversal"?
When the encoder jitters a minuscule amount while across or crossing a boundary between clear and black you will get one step back, one step forward one of more times. Of course, if it jitters a whole lot, it does not need to be exactly across the boundary between clear and black.

I have seen not detecting all double reversals in a tightly coupled encoder (same shaft) resulting in errors of over 1% and when the coupling is as lose as a belt, I can imagine you will see a whole lot more double reversals. (scoping for double reversals I have seen pulses of less than 0.5 uS on a 10 RPM machine with a 5000 step encoder). Realizing the problem and using a proper quadrature encoder instead of the stupid b=dir, a=clock took care of the problem. Since the b=dir, a=clock does take care of 75% of reversals, the error would be 4% or greater shaft coupled. In an 'a only' configuration like what the OP has, each and every double reversal will add to his count.

Erik