Just went through this thread, and it looks very familiar to me
I had this same problem a while ago. To sort it out, what you're seeing is lost steps upon reversals. An easy test can be made to prove it. You don't need to home your machine and read the dial indicator after that. Leave the same setup with the dial indicator and use incremental jogging in Mach3. Set the increment to, say 0.005", then jog one increment towards the dial indicator (remember the reading), then one increment backwards, then again towards. The dial indicator reading theoretically should now show the same as the first time you jogged towards the the clock. In your case I'd bet it would show 0.0015" off.
If this is the case then your loosing steps on reversals. Your stepper driver expects the direction change pulse to appear in a particular correspondence with the step pulse. Some drivers work on the leading edge of the pulse, but most, I think work on the trailing edge (don't quote me on that). For the last case, for example, it means that a pulse is registered be the driver only at the trailing edge. So depending on when the direction pulse occurs relative to the step pulse, you get an axis reverse in the correct moment or not. If the direction pulse is sent "late" then all the steps(actually one) sent till that moment would still move the axis in the previous direction.
Theoretically, changing the active state of the step pulse should solve the problem. But sometimes it doesn't work (as it was in my case). Sometimes, the break out board messes up things. To check this you'd have to bypass the beakout board and feed the signals form the PP directly to the driver. Assuming your PP output voltage is sufficiently high for the driver.
If you want to learn more on the subject and my path in solving this problem, follow the thread "HELP! Accumulative moves" on Mach3 Yahoo Group, at http://groups.yahoo.com/group/mach1mach2cnc/message/102773
And also the thread "HELP! Accumulative moves - Good News" athttp://groups.yahoo.com/group/mach1mach2cnc/message/102937