I believe that a well factored documentation context graph is superior way to model that context, and that there are many advantages to storing that context right in the repo. I think that specs for features should be a a PRD-ERD hybrid that should be deconstructed into a well-factored documentation context graph that can help orient the agents and the humans working with the system. Basically, the 'spec' part of a documentation context graph should be at a PRD and light ERD hybrid level.