Can also be reproduced with this minimal working example:

[general]
# > Verbosity level defining the printout: Options: -1 = super-silent, 0 = silent, 1 = default, 2 = verbose, 3 = debug
verbosity = 1
# > Number of parallel processes to use. Corresponds to the number of physical CPU cores used. Options: <int>
parallel = 4
max_cycles = 200
# > Number of molecules to generate. Options: <int>
num_molecules = 4
# > Do post-processing after the optimization with another engine (e.g., `orca`). Default: false. Options: <bool>
postprocess = true
# > Switch molecule structure XYZ writing on and off. Default: true. Options: <bool>
write_xyz = true

[generate]
# > Minimum number of atoms in the generated molecule. Options: <int>
min_num_atoms = 3
# > Maximum number of atoms in the generated molecule. Options: <int>
max_num_atoms = 5

[refine]
engine = "xtb"
hlgap = 0.5
ncores = 2

[postprocess]
engine = "orca"
optimize = true
opt_cycles = 3
ncores = 4

@lmseidler, could you look into this? According to htop in the initial seconds, the refinement step (here with xtb) does still do something, but then no process is allowed to start anymore. Probably, postprocess wants to run with 4 cores, but only 2 are available. As it happened with both turbomole and orca, I'm quite sure that it's not a problem with the QM programs.
Could it be a problem with the logic in setup_blocks (https://github.com/grimme-lab/MindlessGen/blob/main/src/mindlessgen/prog/parallel.py#L44)?

Please try with the parallelization-setup branch whether this works now. I made a small mistake in the setup logic.

Please try with the parallelization-setup branch whether this works now. I made a small mistake in the setup logic.

If that doesn't complicate things too much, please just open a PR so that I can checkout the branch from there and we can directly merge what was tested.

Found the branch and tried it. Seems to work, great!

If you open the PR, we can quickly merge this, IMHO. Could you add maybe two sentences on what was wrong before?