There's two forms of interpolation going on here that I'm not sure you or Dr Dueschle are using. We interpolate a "band of sight" of single a degree for our azithmual projection, but uniquely we also rotate the DEM elevations around all the observers rather than the observer around to see all the elevations.
The effects of the first can be lessened by lowering the band of sight such that we only process half a degree at a time so that we make sure we get more coverage further away. We plan on running some more experiments by rotating to cover more points.
The algorithm is already fairly expensive to run against the whole world so we weren't particularly interested in that level of coverage for the full earth.
For total viewshed area, our algorithm comes in at roughly a percent or so difference which was what we used as our benchmark for correctness.
All this to say, no, we don't think you both are wrong, we've been looking at making ours more accurate. At a world scale that's quite computationally expensive, so we didn't use that methodology for our initial launch. We see our results as validation of yours, not as something we've disproved.
Edit: grammar
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.
I'm afraid I don't have a good answer. I'm sure with future runs will get closer to you and udeuschle.de
I thought of you when we saw Colombia appear so high up in the list, I remembered that's something you'd found too.