We are now at a point where the tech can help with both of those, today. You can have a cc session "in a loop" going through your docs / code and try to do x and y, and if it gets stuck, that's a pretty good signal that something sucks there. At least you can get a heatmap of what works ootb, and what needs more eyes.

Both questions are getting scary good answers from the latest models. Yes, I tried, on a large proprietary code base which shouldn’t be included in any training set.