Is the image with backlash enabled? Looks like it with the extended parallel segment at turn point. Enabling backlash extends object size as it adds a extra move on direction change. So assuming this is with BS enabled, your problem is a different but we have no image of that problem.
Also make sure not to test first layer. Here different bed height can make the first layer wider or narrower, only the following layer are relevant to distortion.
What other reason for distortion is steps per mm being wrong. Especially if belt is under high tension, it could be different from theoretical value computed from belt pitch and pulley tooth. Not much but measurable on larger objects. In that case the error is proportional, so make 3 rectangular qubes, e.g. 2cm, 4cm and 6cm and see if they grow the same percentage.
