I look forward to seeing what your experiments produce! 
Ah yes–I don’t think that I’ve actually looked at what the Shader Generator does. ^^;
(To be honest, I’m somewhat new at working with shaders; I have some (albeit limited) understanding of their workings, but have very little practical experience in working with them.)
Hmm… That seems as though it could work–I’d likely be worried about all of those texture lookups (if I’m understanding you correctly), but then I imagine that supersampling does something very similar anyway. At the least it does seem worth a shot.
A thought: could it be improved by implementing the run-detection as a binary search? instead of checking each sample point in a given direction, test the maximum distance, and if that’s clear, test halfway to it, and so on. That might allow you to test longer runs than an iterative approach–but I don’t know whether it would be feasible in a shader.
Oh yes, it would be a completely different approach, with a thoroughly different set of components and challenges!