[rish-16]

docs/inference.md says molecule structure prediction perceives bonds with Open Babel when bond_sidecar_path=null, but only ZatomContigJsonDataset (validation) did this. zatom2 design loads specs through RFD3's ContigJsonDataset, so without a sidecar the model got an almost empty token_bonds graph and generated molecules with the right formula but the wrong connectivity.

• Perceive bonds from preserved native coordinates in to_pipeline_input, which every inference path goes through.
• Move _can_perceive_ligand_bonds to bond_sidecar.py and share one helper with ZatomContigJsonDataset (no double perception; sidecars still take precedence).
• Add test_contig_inference_perceives_bonds_without_sidecar, covering the CLI loading path.

[amorehead]

LGTM. Thanks for your contribution!

[Copilot]

🟡 Changes recommended

Sidecar-backed ligand inputs now require Open Babel before their precomputed bonds can be loaded, breaking those inputs when it is unavailable.

1 open issue

• Bypass Open Babel perception when sidecar bonds are configured · New

What changed in this PR

This PR addresses missing molecular connectivity in zatom2 design by moving bond perception into the shared inference-input conversion path.

Changes:

• Perceive ligand bonds using preserved native coordinates before inference transforms.
• Share perception logic with the validation dataset and avoid repeated perception.
• Add regression coverage for the CLI’s ContigJsonDataset loading path.

File	Description
tests/​test_bond_sidecar.py	Tests inference bond perception without a sidecar.
src/​nucleus/​utils/​datasets.py	Reuses shared perception logic and skips it when already completed.
src/​nucleus/​utils/​bond_sidecar.py	Adds the shared native-coordinate bond-perception helper.
src/​nucleus/​inference_input_parsing.py	Invokes bond perception during inference-input conversion.

_{Review effort: Balanced}