refactor the fact-generation code to be less messy, use fewer MIR walks

[amandasystems]

I propose something like this:

Green is new, red is removed, bold is a module, and regular font is a fact.

Facts are grouped based on the inputs they require to compute.