Regarding purges, software updates and overwrites, these images are no different than the thumbnails of arbitrary sizes we currently have. If there's an issue with any of these aspects, it's already affecting us and it's unrelated to the mipmapping discussion.
* Currently, JPEG images are sharpened. Does this affect how the
end-result looks like if they are sharpened twice? Do you expect other
quality-issues, for example for GIF or PNG?
Since GIF and PNG use lossless compression, loss of quality for those formats will be even less visible than JPG. That's why I've focused my tests on JPGs so far, it's the worst case scenario. If JPG visual quality can be maintained with this technique despite the lossy nature of the JPG compression, GIF and PNG won't be a problem.
Is there a comparison of
applying down-scaling twice with doing it just once?
I'd like the verify the visual impact of downscaling every step first, because it offers the biggest performance gain. We can always look at the alternative of having only one big as plan B.
I've just sent a survey to wikitech-l [1] to gather some human results and verify if there's any statistically difference in perception between JPGs that have been generated through the existing image scaler and images that have been resized 4-5 times.
[1]
http://lists.wikimedia.org/pipermail/wikitech-l/2014-May/076286.html