I believe that for many “large” scientific projects having file-dependent compilation flags (and maybe other similar fine tuning functionality) is crucial, mainly when dealing with the inevitable amount of legacy stuff (that may well not reach the API, but still has to be compiled). As an example see my comment on a related post:
More generally both CMake and cargo (which is the main inspiration for fpm, as I understand) provide easy directory-based differentiation of the build options. The former via the pervasive nested CMakeLists.txt + add_subdirectory pattern, the latter I don’t know well, but I see it’s extensively used in DFTWorks. But for what I got skimming through several disseminated discussions (github issues mostly I think) this is not planned for fpm, so that could be a nontrivial barrier for very large and especially heterogeneous projects…