Virtual sites not right for terminal NH2-group

GROMACS version: 2018
GROMACS modification: No
Hi dynamicists,

I am suspecting that pdb2gmx are making the wrong virtual sites constructions for neutral N-termini. A warning is given by grompp:

Program: gmx grompp, version 2018
Source file: src/gromacs/mdlib/calc_verletbuf.cpp (line 253)

Fatal error:
In molecule type ‘Protein’ Virtual site 3fad construction involves atom 5,
which is a virtual site of equal or high complexity. This is not supported.

This does not happen with a charged terminal amine. I have inspected the topology and seen that the two hydrogens indeed have 3fad vsite constructions that rely on a hydrogen constructed as an 4fdn vsite, which is disallowed.

What also puzzles me is that the two amine hydrogens are constructed as 3fad in the first place. They are constructed as HT1-N-CA-HB1 and HT2-N-CA-HB1, respectively, in an N-terminal histidine, but I assume it would be the same in most aa types. If I understand correctly, this would connect the side-chain rotation to the geometry of the NH2, which seems odd to me. Is this really the right virtual site construction for terminal NH2-groups?

PS I tried generating the topology with pdb2gmx using 2021.2, but the vsite constructions look the same.

Kind regards,