Parallel MD

[19hello]

Reminder

• I have read AGENTS.md and docs/developers_guide/agent_governance.md.
• I have linked an issue or explained why this PR does not need one.
• I have added adequate unit tests and/or case tests, or explained why not.
• I have listed the exact verification commands run and their results.
• I have described user-visible behavior changes, including INPUT parameter changes.
• I have explained core-module impact for ESolver, HSolver, ElecState, Hamilt, Operator, Psi, or other source/ changes.
• I have requested any needed governance exception below.

Linked Issue

N/A. This is PR3 of the parallel-MD series, following the merged distributed MDCell infrastructure from #7768.

Unit Tests and/or Case Tests for my changes

• Commands run:
  • cmake --build /tmp/abacus-pr3-full -j4 --target abacus_basic_para MODULE_MD_LJ_pot MODULE_IO_input_test_para
  • env LD_PRELOAD=/usr/lib/x86_64-linux-gnu/openblas-openmp/libopenblas.so.0 ctest --test-dir /tmp/abacus-pr3-full -R '^(MODULE_IO_input_test_para|MODULE_MD_LJ_pot)$' --output-on-failure
  • env OMP_NUM_THREADS=1 LD_PRELOAD=/usr/lib/x86_64-linux-gnu/openblas-openmp/libopenblas.so.0 bash ../integrate/Autotest.sh -a /tmp/abacus-pr3- full/abacus_basic_para -n 4 -r '^(01_LJ_Anderson|04_LJ_Langevin)$'
  • python3 tools/03_code_analysis/agent_governance_check.py --staged
• Result summary:
  • MODULE_IO_input_test_para and MODULE_MD_LJ_pot passed.
  • The two 4-rank random-MD cases passed all energy, force, and stress comparisons after updating their references.
  • A 4-rank native MDCell LJ run verified that md_out_force=0, md_dumpfreq=0, and md_restartfreq=0 suppress TOTAL-FORCE, MD_dump,
    STRU_MD_*, and Restart_md.txt without affecting MD execution.
  • Governance check reported no findings for the final reference update.
• Checks not run, with reason:
  • The complete MD C++ unit-test set and full integration suite were not rerun after the final reference-only commit. The final commit changes only
    the two validated reference files.

What's changed?

• Enable native distributed MDCell runtime workflows for MD-capable LJ, DP, and NEP solvers without constructing a global UnitCell for the native
  path.
• Keep atom ownership, migration, neighbor construction, force evaluation, MD restart, and MD output distributed by MPI rank.
• Use 64-bit atom counters in the distributed MDCell path.
• Stream MDCell TOTAL-FORCE output to rank 0 without allocating a global force array.
• Add md_out_force (default true) to control the MDCell TOTAL-FORCE table in running_md.log.
• Allow md_dumpfreq=0 to disable MD_dump, and md_restartfreq=0 to disable STRU_MD_* and Restart_md.txt.
• Initialize an independent MD random-number stream on each MPI rank. Anderson and Langevin 4-rank references are updated accordingly.

Governance Notes

• INPUT/docs changes:
  • Added and documented md_out_force.
  • Documented zero-frequency behavior for md_dumpfreq and md_restartfreq.
  • Updated md_seed documentation to describe rank-local seeding.
• Core module impact:
  • source_cell: adds and maintains distributed MDCell state, STRU reading, migration, and distributed output.
  • source_md: drives MD directly on MDCell; MD depends on IO only for existing parameter and output facilities.
  • source_esolver: LJ, DP, and NEP provide native MDCell force/stress/cutoff support.
  • No HSolver, ElecState, Hamilt, Operator, or Psi workflow changes.
• Exceptions requested:
  • None. The remaining GlobalV::ofs_running use in the MD-to-IO force-output call follows the existing output API; the relocation reduces rather
    than increases the PR-level global dependency count.