I agree, but it’s painful when it goes up.
For me, this is more of an issue of which headers to include than a deeper backward compatibility issue. We haven’t been very good about backward compatibility for the math lib, to say the least. We’ve been much better about the Stan language itself.
I think those timings are going to vary significantly based on disk speed since you’re reading in much more code. It’s all just reading and dismissing what you don’t use, but our headers include a gazillion other headers, and that’s all file handles opening and closing and reads happening.