SOF-8005: parse NWChem initial and final geometry

[VsevolodX]

This PR (express)

NwchemParser implemented none of IonicDataMixin's four geometry methods, whose bodies are bare pass. So lattice_basis_to_poscar raised on None, and every NWChem job ever run silently failed to publish its structures — extract_structures() runs app-agnostically for every modeling unit, and extract_property swallows the exception into a log.error on the compute node. That predates the relaxation work entirely.

One extractor keyed on the Output coordinates block, plus four thin delegates. Three things worth a reviewer's attention:

• The block header carries the units the input declared, and they are not always angstrom. tests/fixtures/nwchem/test-001 is a geometry units au run that prints no angstrom block at all — the plan's instruction to key on the literal Output coordinates in angstroms would have returned None for it, leaving that silent failure open for every au-geometry job. The units word is read rather than assumed. An angstroms block is taken verbatim rather than rescaled through two Bohr radii that disagree in the last digits.
• The cell is not ours to invent. made's calculate_padded_cell_simple_cubic, the same convention every non-periodic material on the platform gets. Initial and final therefore get different edges, as made would give them: 5.72171236 Å vs 3.05045924 Å for the H2O fixture. The basis takes its edge from _lattice_vectors so it is centered in the box that ships with it.
• BASIS in tests/fixtures/nwchem/references.py was wrong and unused. It held test-001's printed a.u. geometry tagged "units": "angstrom" — added in the original NWChem PR and never referenced by any test, so nothing caught it. Corrected to that geometry converted and centered, and now actually asserted.

Tests: 22 passing (unit + integration), including a regression case for a numeric row that fits the column shape but has no element symbol — it is skipped rather than raising, because rupy would swallow the exception into a silently missing final_structure.

Pre-existing on main, untouched: 11 failures in tests/integration/parsers/test_structure.py, tests/unit/parsers/test_structure_parser.py and tests/integration/properties/test_material.py, all espresso structure parsing. Confirmed by stashing every change and re-running.

---

Part of SOF-8005 — [Standata] Relaxation SWF/WF for NWChem. One ticket, five repos; this is one of five PRs that land together.

The goal, stated once: relax a molecule the same way we relax a material.

total_energy computed HOMO/LUMO on whatever geometry was uploaded — nothing in the NWChem workflows relaxed anything, and get_relaxation_by_application("nwchem") returned {}, so Toggle Relaxation silently no-opped. These PRs close the four gaps between NWChem and what VASP/Espresso already do.

#	Capability	Repo
1	A relaxation subworkflow exists and Toggle Relaxation resolves it	standata
2	The next unit computes on the relaxed geometry	wode + esse + standata
3	The relaxed structure comes back as a Material	express
4	A test proves it end to end	web-app

Evidence

Acceptance test job-nwchem-relaxation-total-energy.feature passes against a live cluster — takes the unrelaxed H2O.poscar, toggles relaxation in the real UI, runs one job:

property	value	reference (NIST CCCBDB, B3LYP/6-31G)
homo_energy	−7.880687 eV	−7.880
lumo_energy	1.438531 eV	1.439

Tolerance is 0.02 eV, not the 0.1 eV the other nwchem features use: relaxed and unrelaxed HOMO differ by only 0.053 eV, so 0.1 would pass on an unrelaxed molecule — precisely what this test exists to catch. LUMO differs by 0.341 eV and is the discriminating assertion.

The sharpest evidence is not in the assertions. The relaxation unit reported total_energy = -2078.5728586862333 and the total-energy unit -2078.5728586879750 — agreeing to 2e-9 eV. The second unit reproduced the first's final energy, which it can only do by computing on the relaxed geometry; a start-from-scratch unit would have recomputed the original upload.

Manual checks

• Rendered input of the total-energy unit says restart nwchem with no geometry block — read from a real job at job.workflow.subworkflows[].units[].input[0].rendered.
• [~] The relaxation converged — HOMO/LUMO reproduce published equilibrium values to ~1 meV, which a truncated optimization would not. The literal Optimization converged string was not read; job files are in object storage behind the authenticated API.
• A relaxed Material appears with a resolvable materialId — still absent, and deploying express will not fully fix it. is_non_periodic appears nowhere in rupy and Material.__init__ defaults it to False, so the relaxed molecule builds through CrystalParser: cubic lattice type yes, isNonPeriodic no. Timur asked for exactly this on this exact file in express#55 (2020-12-08): "Are we able to retain the fact that this is an isolated molecule, rather than a periodic system…". Needs a rupy change; scoped out here and filed as a follow-up.
• The generated box — 3.05045924 Å for relaxed water, 5.72171236 Å unrelaxed. Correct by made's convention; periodic images essentially touch, so it must be re-boxed before any periodic reuse.
• Toggle Relaxation stamps a sane version/build — {nwchem, 7.0.2, GNU, isDefaultVersion: true}.
• npm and python standata both ship it — both lookups return Geometry Relaxation.
• Frequency (nwchem) with no relaxation renders start nwchem, byte-identically to before.

Review

Two tb-reviewer passes. The first found two blockers — local file: tarball pins in web-app, and a jobHasParent || half that would have rendered a deck NWChem cannot run — both fixed. The second, grounded against the precedent corpus (12,340 comments / 73 repos), found that Toggle Relaxation had become reachable on Frequency (nwchem) while that template still declared its own geometry: it would have relaxed, discarded the result, and reported frequencies for the original molecule, silently. Fixed here.

Known and deliberately not fixed

• rupy never passes is_non_periodic — above; needs a rupy change.
• isRelaxed will read False on every NWChem relaxation once express is deployed. _is_relaxed() derives it from pressure and atomic_forces, and NwchemParser implements neither. Unreachable today because extraction fails first.
• Relaxation resolves by application name alone (wode Workflow.ts:352), not by what the workflow computes. wode#100 already pointed this way.

🤖 Generated with Claude Code