Performance: the momentum residual's face kernels compile to specialised straight-line code again (forced inlining, constant directions, one transverse-stencil selector); ComputeRHS is now faster than before the rhs.c rewrite, with byte-identical results to it.