If we can ensure this is the case, it should help our incremental inference perform better.

One issue here is that low-durability files (in your workspace) can shadow high-durability files (in the stdlib), and in principle this could impact the type inference in the stdlib. This means that any resolve_module is (I think) currently a low-durability query, because it will always access some low-durability import search paths.

If it helps performance enough, we could decide that we never want type inference within the stdlib to respect shadowed stdlib modules, so that we can keep type inference of the stdlib all high-durability, and not need to revalidate it in depth on incremental changes to user code.

This maybe should go hand-in-hand with always warning or erroring if a user module does shadow a stdlib module.