Technically, it now works by getting the mouse motion in 3D (just like Translation/Grab), projecting it on the vertical view axis, using the vector length as the shrink/fatten factor. If the motion is downward (on the screen), shrink, if it is upward, fatten.
In layman terms: move up to fatten, down to shrink and it adapts to the viewport zoom, like Translation.
I changed to snapgrid factors to match those of Translation to, so it really acts like grabbing.
bc9291d806