feat: add DPA-style NequIP and SevenNet energy-plus-property multitask

[anyangml]

Summary

• Reuse the MACE DPA-style split for NequIP and SevenNet: shared descriptor, original energy head as a DeePMD fitting (nequip_ener / sevennet_ener), PropertyFittingNet on last-layer 0e, two datasets via model_dict / data_dict.
• NequIP keeps output_hidden_to_scalar plus stored e0. SevenNet keeps the native readout and species-wise rescale, packing full last-layer features beside property 0e.
• Training still skips a second vacuum-energy fit; finetune change-by-statistic is a residual out_bias. model.type: nequip / LAMMPS and the MACE path are unchanged.

Test plan

• pytest tests/test_nequip_ener.py tests/test_sevennet_ener.py tests/test_examples.py tests/test_nequip_descriptor.py tests/test_sevennet_descriptor.py tests/test_mace_ener.py -m "not slow"
• Confirm NequIP energy/force match native NequipModel on the serialized artifact
• Confirm SevenNet energy matches the original head on the same descriptor graph
• Confirm property 0e is unchanged and share_params level 0 aliases the backbone only
• Confirm checkpoints restore after the native pickle/artifact is moved

[codecov]

Codecov Report

❌ Patch coverage is 93.12796% with 29 lines in your changes missing coverage. Please review.
✅ Project coverage is 86.81%. Comparing base (adda02d) to head (be9cfbd).

Files with missing lines	Patch %	Lines
deepmd_gnn/nequip_ener.py	90.00%	16 Missing ⚠️
deepmd_gnn/sevennet_ener.py	93.23%	9 Missing ⚠️
deepmd_gnn/pt.py	93.75%	4 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #173      +/-   ##
==========================================
+ Coverage   85.95%   86.81%   +0.85%     
==========================================
  Files          22       24       +2     
  Lines        2941     3352     +411     
==========================================
+ Hits         2528     2910     +382     
- Misses        413      442      +29     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[njzjz-bot]

I found one numerical-correctness blocker in the new set-by-statistic handling for native GNN energy heads. The rest of the NequIP/SevenNet split, checkpoint self-containment, descriptor sharing, native-head parity tests, and current exact-head CI look consistent; the concrete issue is attached inline.

Agent: ChatGPT
Model: GPT-5.6 Sol
GitHub account: njzjz-bot
Reviewed head: bd1cc9c
Trigger: scheduled all-PR monitoring

[njzjz-bot]

Re-reviewed the new head. The previous scalar-SevenNet-shift blocker is fixed: the native scalar Rescale.shift is now preserved, while the type-resolved set-by-statistic correction is stored in DeePMD out_bias. The new scalar-shift regression directly checks that the native scalar is unchanged and that the per-type correction remains distinct, so the earlier silent arithmetic-mean loss of the energy baseline is no longer present. I did not find a new high-confidence blocker in this delta.

I am not approving yet because exact-head validation is still incomplete: Test CUDA and pre-commit.ci are green, while Test Python/package, PyPI build/release, and Read the Docs are still pending/in progress. I also attempted to resolve the now-outdated blocking inline thread, but GitHub rejected that thread-state mutation with Resource not accessible by integration; the old thread may therefore remain visually unresolved even though the code issue is fixed.

Agent: ChatGPT
Model: GPT-5.6 Sol
GitHub account: njzjz-bot
Reviewed head: be9cfbd
Trigger: scheduled all-PR monitoring

[njzjz-bot]

Exact-head validation is now complete. Test CUDA, Test Python package, PyPI build/release, pre-commit.ci, and Read the Docs all pass on be9cfbd8824773176876202ec14f057bd37b35fc. The previously reported SevenNet scalar-shift numerical-correctness blocker is fixed on this head and covered by the added regression; no new high-confidence blocker emerged in the completed validation.

Agent: ChatGPT | Model: GPT-5.6 Sol | GitHub: njzjz-bot | Reviewed head: be9cfbd | Trigger: scheduled all-PR monitoring