The idea is that since we're racing against the electron beam, every code is dependent on it and failing so will result in screen tearing
(think of it as rolling stutter but between frames). If the frame didn't finish to process before the electron beam reached the point, so the screen partially unpdates. Now, this usually doesn't matter since most registers are h-/v-/f-blank locked but this still can happen with HDMA where its table can be located in RAM and that can be written everytime.
The effect doesn't happen in SMW hacking because there is so much NMI time left, screen tearing generally won't appear or isn't noticable and the HDMA effects are generally simple (there also is the fact that getting some time). Even if that happens, this generally isn't appearant on most HDMA effects (such as gradients or parallax in general) but certain other effects such as windowing of fixed shapes or fast changes are noticable.
Note that this doesn't apply to changes in the scanline but changes to the HDMA table in general (including the value).
Now you understand what screen tearing is but it also doesn't seem to be related to your problem, though.
Okay, my layout looks ugly.