See e.g. http://download.rockbox.org/daily/manual/rockbox-sansaclipplus.pdf starting at page 2: Some glyphs render with a higher/lower baseline than the others in the same word. This regressed with the imagesmooth.c changes from http://code.google.com/p/sumatrapdf/source/detail?r=2614
This should come to me I think. The proposed patch actually puts back code that I removed to ensure that rotated pages looked identical when rotated back.
Created attachment 7237 [details] patch.txt Fixed version; I was miscalculating the y sub pixel position. Previously I was calculating it as if the top left of the scaled image was going to be displayed at the given (x,y) position, not the bottom left.
Closing, as Tor has pushed this to the repo. Thanks.