GoMartini topology produced using Martinize2 doesn't run on GROMACS

[OLaprevote]

Using this command:

gromit/martinate.sh -f myprotein.pdb -ffdir martini_v300 \
    -martinize martinize2  -cg martini3001 --martinize-p=backbone --martinize-govs-include \
    --insane-d=8 -stop CG

I believe this should normally produce a GoMartini model given martinize2 documentation about -govs-include. The molecule_0.itp file produced indeed includes virtual Go sites:

[ atoms ]
…
 95 Q5            42 ALA BB   95   -1
 96 TC3           42 ALA SC1  96  0.0
 97 molecule_0_1   1 ASP CA   97    0
 98 molecule_0_2   2 ALA CA   98    0
…
[ virtual_sitesn ]
; Virtual go site
 97 1   1
 98 1   3
 99 1   5
…

However the first energy minimization (before solvation by insane.py) returns an error (in 01-TOPOLOGY-CG.log):

ERROR 1 [file molecule_0.itp, line 111]:
Atomtype molecule_0_1 not found

I believe that entering the --martinize-govs-include argument should return a warning or an error until Go̅ model are up and running with martinize2.

The subsidiary question is: using martinize2 and martini3 forcefield in martinate.sh, how should I make a Go̅-like model of my protein? What am I missing right now to get a valid topology for GROMACS? Tutorials are scarce and dedicated to martini2, but given this paper I believe it is already doable.