I don't appreciate how the author dismisses precompiled headers, then introduces a header-to-module mapping tool suffers from all of the same limitations. It seems to me that using a tool assist with managing header dependencies in conjunction with PCH will solve 80-90% of the problems this module mechanism attempts to solve.
It's also not clear to me how the modularized code gets transformed to a format that the native linker cannot understand. Systems programmers are not going to use modules if it means using LLVM's linker.