G'day (again)!
As I stated at dslm4515/CMLFS#41, back in September 2021 --- Lord, that's a damn amount of time! ---, citing Gentoo developer Agostino Sarubbo's blog post, using more than one job per processor isn't ideal --- but mussel is doing not just two jobs, but three jobs per processor core ---, because it makes the build slower since the processor suffers from such overload.
I've got the data: I've ran current main branch (10a8a05) mussel along with a quick-n-dirty hack that defines $JOBS just as getnproc(), without triplicating it, for building a toolchain with Linux kernel headers (-l), OpenMP (-o) and pkg-config (-k) support, and the results were a difference of around 15 minutes --- w/o the fix: real 1h11m28,67s, with the fix: real 56m49,27s ---, plus the fact that the machine was also capable of doing another tasks because there wasn't such overload.
I don't know what was the reason for doing it before, but, at least in the most cases, it is, indeed, the most efficient way.