Good idea, I'll add it to the FAQ later today. Under a section of "why don't these results match the other tools". The projection error is separate as you mentioned.

The error I've experienced hunting bugs tends to be within about .5-2%. That's a vibe, not an empirical "I've calculated the error to be 1.5%". We definitely expect that bound to tighten as we get access to more computational resources.

I do not think this is numerical however. I think it's more directly related to rasterization, interpolation, and not enough angle coverage. We have fairly good numerical and viewshed tests to double check we don't have weirdness going on there.