Backlash compensation - strange behaviour on circular shapes

Backlash compensation's implementation is simpler than you think (you're over thinking it)

It is merely an additional movement added to the beginning of the 1st move after any direction reversal. no more no less. It is made at the fastest speed possible (limited by the stepgen's max velocity and acceleration. How quickly it may be taken up will depend on the headroom given between those max speeds and the commanded feed rate of the move. Once the move is applied all movement is as commanded until the next direction reversal of that axis.

So for milling a circle, there will only be at the most 4 backlash compensating moves one for each of the two X and Y reversals. If everything is perfect you get a smooth circle, but usually you will end up with a small imperfection at each of the 4 reversal points. If you have used too much backlash compensation you will actually get a little step inward at each of the reversal points. If you arn't using enough backlash comp, the effect is difficult do describe, sort of a flat spot and slight distortion of the arc, and isn't as noticeable.

The bigger problem is that cutting forces can work against you taking up the backlash without or before the backlash move, resulting in an obvious step when the backlash move is made resulting in all sorts of ugliness.

Maybe you're right, it sounds reasonable.
But... how can this explain my almost perfect 10 mm cylindrical-pockets* without compensation**, versus my buggy 10.2 mm pockets with compensation active, on a soft material, like the MDF is?
As you pointed out, the micro-steps at reversal points should be directed inward, not outward. The resulting shape, in the worst case, should result ugly but smaller, not bigger. I'm missing something?

(*) In one case I have registered 10.00 mm, that is the max precision of my caliper.
(**) 0.2 mm of compensation applied on Y and X axis both .

What size tool were you using? Describe your machine and spindle?