Yeah! I loved learning this (some time ago) too! Wonder whether there's any cool filtering tricks that can give you something similar but also filtered result. Ie an 'anti-aliased' version - although interestingly the method in the article mentions the reversibility of the whole-pixel-shears approach, which is a fun exception to what 'aliasing' usually means: worse recovery of original signal.
Yeah, you resample fractional pixel values in the shears, ideally with a sinc() kernel. This has the virtue that the resampling weights are identical across lines in the shear, which allows a very efficient SIMD implementation. This is how we do CPU-based image rotations!
