There are lots of ways to skin this cat. C++ has what they call a "large installed base" -- i.e., literally hundreds of billions of lines of code in active use. So, the solution for C++ is not what we might have done for a new language. It had to be reasonably easy to move an existing project to modules without breaking everybody that might depend on it, and without needing everything you depend on to be moved already. Everybody has their own upgrade schedule, many of which must be "never".
So, it looks odd compared to a green-field approach, but it works and is not a burden to maintain, which is what matters.
C++ headers are a total mess, and modules are the cleanup effort.
It sort of makes sense. The calling program specifies `foo.bmi`, but compiling for ARM might look in "armv5/foo.bmi" while targeting x64 might look in "x64/foo.bmi", and that doesn't even take into account that there might be different debug and release versions.
For your example, you could make a 'foo' module that pulls in an implementation from a traditional C++ source file, "armv5/foo_impl.h" or "x64/foo_impl.h", using preprocessor #if/#ifdef to choose the right one. Or you could even use a module backend, pulling in the implementation from either "foo_armv5.bmi" or "foo_x64.bmi".