Ah, I see the issue now. I was looking at dH/dlambda, as in the title of your post, that is present. But the foreign lambda terms are all zero and that is what gmx bar uses.
I filed an issue: Foreign Hamiltonian differences for perturbed masses are zero (#5195) · Issues · GROMACS / GROMACS · GitLab
And I uploaded a fix: Draft: Add effect of perturbed masses to foreign Hamiltonian differences (!4661) · Merge requests · GROMACS / GROMACS · GitLab
Thanks for reporting this.