NepTrain2.2.0 PythonAPI and ABACUS labels | real Si2/Si8 examples

1. bash run.sh : print included real numerical reference and redraw3figures; no DFT/training.
2. sbatch job.slurm : in a fresh extracted directory, run52 Python-interface examples, two real8MPI ABACUS PW SCFs, independent native-output analysis, canonicalF export and figures.
3. In an allocated8core-or-larger task, bash api.sh runs only Python examples. Do not rerun full.sh over completed work/abacus or api.sh over completed work/api. Use a new extracted working copy for a new run; outputs are deliberately protected.

The CPU allocation requests56physical cores under the cluster whole-node rule; this teaching example actually uses8physicalcores/8MPI/OMP1. It is not a56core benchmark. No GPU is used.

The full runner loads gpumdkit/1.5.8 (Python3.11/NepTrain2.2.0/ASE3.29.0), and the native wrapper loads ABACUS3.10.1-8173m (InteloneAPI2024.2/IntelMPI2021.13). OfficialASE-ABACUSsource is pinned and process-local; the shared environment is unchanged.

Inputs: data/test.xyz contains8 own earlier realPBE Si reference geometries; abacus_case.py selectsoneSi2/oneSi8 and removes oldcalculatorlabels before nativeSCF. PBE60RyPW/cg/scf_thr1e-8/ka12,12,12; actualKgridswritten bythe interfaceare4,4,4 and3,3,3. No convergence/performance/production-potential claim.

Raw installed NepTrain labels: work/abacus/labels.xyz. Raw ASE energy = nativeF - E_entropy(-TS) (internalE), notF and notE_KS(sigma->0). analyze.py preserves the original and writes analysis/abacus-labels-F.xyz with explicitly labeled nativefreeenergyF and unchangedforces/virial. DPDATA independently reads nativeenergy, forces, virial; the original numericalgates are unchanged. DifferentDFT codes/pseudopotentials/energydefinitions are not mixed into training merely because unitsmatch.

Output: analysis/neptrain-*.csv (5tables), API-RESULTS.json, SUMMARY.json; figures/*.png(3). Data-driven NEP descriptor projection is only a2-Ddisplay; actualFPS usesfull30dimensions. SOAPhas9featuresfor specifiednmax2/lmax2. Constructedperturbations and halfAngstromnegativecontrolare notMDtrajectories. NativeMC-rattle legacyhelpersexistintheinstalledsource butare notdefaultCLIperturbation. Inputrestartdemo usesanactualnativecheckpoint anddoesnotclaimnewresumetraining.

installed-symbols.csv indexesall94source-level definitions (including3underscorehelpersand1asyncfunctionthatwerenotintheearlier90synchronousnonunderscorecount). Internalhelpers andclassmethodsarecataloguedwithoutclaimingeachisastablepublicAPI,externalPBS/cloudexecution,orallparameterbranchesarenumericallytested. ActualCLI,workflow,training,MD,VASP,andcoldresumeexamplesareintheearlierNepTraintopic; theyarenotrepeatedhere.

Native wrapper is invoked explicitly through bash, so the archive also works when the extractor does not preserve executable permission bits.
