I've found that SDD is actually what you need to be able to work with code bases when they go above around 100 000 lines of code. It's what unlocked getting LLMs to work well with large codebases for me.

I still don't get it, can you clarify? It's not the research phase that I'm disputing. Clearly for a large codebase, you need some good way to take all that information (code, product knowledge) and distill it down to something that can fit in the context, ready for implementation. And it's that research that is going to get harder the bigger the codebase. (My current experience is with a repo around 1.5 million lines.) I'm saying that the output of that research, in my experience, doesn't need to be anything like the detail of an exact spec. It can be a sort of one-to-two-pager Markdown doc, at most – and any further detail is much more ergonomic for me to iterate over in the form of code.