DFT-D4 host integration examples

1. Inspect the fixed XYZ geometries in inputs/. These are not optimized structures.
2. Replot the actual cluster reference data: bash view.sh
3. Prepare new host cases: python3 prepare.py
4. Enter one cases/ENGINE-MODEL-MODE directory and submit: sbatch job.slurm
   Choose only a partition/node permitted for your account. CP2K and ORCA here
   use one CPU process; the allocation itself is not a GPU performance claim.
5. Calculate standalone D4 for the supplied models: python3 d4.py
6. Compare your completed paired cases:
   bash view.sh --cases cases --d4 d4-results --output my-analysis --partial

CP2K: cp2k/2026.2-p100 or -v100; GNU, MPICH5.0.1, CUDA12.9,
       libdftd4 4.2.0. PBE/DZVP-MOLOPT-SR-GTH with GTH-PBE potentials.
ORCA:  orca/6.1.1 official precompiled program, OpenMPI4.1.8 runtime;
       PBE/def2-SVP. Its native D4 output reports3.4.0.
VASP:  vasp/6.5.1-p100, IVDW=0, plain PBE. Only the external periodic D4
       single-point post-addition is demonstrated. Obtain licensed O,H PAW
       potentials yourself; POTCAR_ROOT may select your licensed PBE directory.
       Raw licensed VASP files are not in this archive.

Never add external D4 again to a total energy which already contains D4.
Do not use post-added energies as native dispersion-corrected optimization/MD.
All energy differences compare identical host/basis/geometry/charge settings.
Do not compare CP2K pseudopotential absolute energies with ORCA all-electron
energies. The data qualify the dispersion correction bookkeeping, not total
DFT force/grid convergence, binding energies, or ionization energies.
