Fortunately most layouts tend to mainly focus on changing the main keys, which makes it a bit easier. For those cases you can easily compare Layout1/ANSI vs Layout2/ANSI, and then separately Layout1/ANSI vs Layout1/Ergodox, which effectively separates out the letter arrangement from the hardware form factor. Changing extra things like numbers etc shouldn't have a huge difference since they are relatively uncommon, although conceivably you might have a heavily numeric-based corpus. Even then, changes to these keys can still be modelled, so I wouldn't expect a need to handle these cases differently.

I think the much trickier issue is how to handle things like modifier changes, the various thumb-key combination, etc. And even more complicatedly: home row modifiers, dual-role keys etc. These more advanced features are harder to model, so analyzers tend to conveniently ignore them for the most part. Most of the effort goes on thinking about finger usage, but thumbs are somewhat overlooked. It's certainly an area that needs more thinking about, especially now that ergo boards with lots of thumb keys are becoming more common.

I guess ideally we'd need some sort of generic model of the human hand which could be applied to all sorts of weird and wonderful layout ideas!