Yes I don't believe the OP has a "25000 steps/rev" motor, it would be very special.
Larger and larger error towards one end sounds like you have just some slightly wrong number somewhere. For example, if you expect your reduction gear is 1:20 but in reality it is 1:21, you would have linearly increasing error ending up at 5%.
Recheck all numbers.
Regarding microstepping, the point is not to magically generate ridiculously high positional accuracy, the point is to drive the motor with sinusoidal waveform instead of square wave and minimize stutter / vibration. The more microsteps, the better it approximates sinusoidal drive. You can't have too much really, use as much as you can implement.
Regarding missed steps, you'll notice if that happens from the jerk and sound it makes. If it is sitting on your lab table and operating smoothly without you noticing any weird, you are not missing steps.
Backlash compensation can be implemented by always stopping the motion in one certain direction. If you need to go to the opposite direction, go half a degree over the destination then come back half a degree.
Steppers are OK for many applications if you know and accept the limitations.