Add a one-process relaxed rotor scan to the ASE adapter

[alongd]

What

Implements job_type='scan' for ASEAdapter: a complete 1D relaxed torsional scan run in a single process, sweeping one torsion in fixed increments and relaxing all other degrees of freedom at each point. Previously ASEAdapter had no scan branch at all — a scan job fell through to the geometry-save else in ase_script.py, so output.yml carried no energies/angles and ARC's YAML scan parser returned (None, None) silently (a quiet empty potential, not an error).

Why one process per rotor, not one job per point

A rotor scan can be decomposed into one constrained optimization per dihedral point (as directed_scan does), but for a machine-learned potential (e.g. UMA) each such job re-loads the model checkpoint to do a few seconds of work — for a large corpus that is hundreds of thousands of tiny jobs dominated by fixed startup. An ESS engine like Gaussian sweeps the whole rotor internally via its scan keyword; this gives the ASE adapter the same capability, so a full rotor is one in-process task instead of ~46 jobs. The model load is then paid once per worker and amortized across every rotor it handles.

The scan loop lives in ase_script.py (which runs in the calculator's own environment); the adapter only writes the input (torsions, scan_res) and parses the output. output.yml carries energies (Hartree) and angles (degrees), read back through the existing yaml parser adapter and rotors_dict['scan_path'] — no change to the statmech/Arkane side.

Design choices

• Bidirectional relaxed walk. The torsion is swept 0→360 and 360→0 from the same starting geometry, keeping the lower energy at each grid point. A single directional walk is path-dependent: if a coupled coordinate flips partway round, that walk ends in a different conformer and its curve is not a torsional potential. Taking the pointwise minimum of both walks recovers the lowest torsional path.
• Geometry propagation is hysteretic — each point starts from the previous relaxed geometry (which is exactly why both directions are walked).
• Per-point constraint. After rotating the increment, the point's actual dihedral is measured and pinned with FixInternals; convergence is judged on the residual force with the constraint still applied (reading opt.converged() after clearing it re-exposes the free torsional force and falsely reports non-convergence at relaxed points).
• Non-converged points are kept, never dropped — a hole in the periodic grid would corrupt the downstream Fourier fit. fmax_worst in the output records the largest residual.
• Rotating group is derived by graph traversal (break the pivot bond, flood-fill from the second pivot), raising on a ring pivot.

Validation

Validated on ethane against an already-validated standalone MLIP scan (UMA uma-s-1p2), same species / torsion / starting geometry:

	this PR (incore ASE scan)	reference
barrier	10.6879 kJ/mol	10.6880 kJ/mol
points × resolution	46 × 8°	46 × 8°
max abs deviation	—	0.028 kJ/mol (mean 0.012)

The two V(φ) profiles are the same curve; the residual is at the floating-point/optimizer-noise level (the reference ran on GPU, this on CPU).

Tests

Adds hermetic unit tests using ASE's EMT calculator (no MLIP needed in CI): rotating-group detection incl. the ring-pivot error, a full 46-point relaxed scan (grid, endpoint duplication, finiteness), output round-tripping through ARC's YAML scan parser, and the single-torsion guard.

Scope

scan only; additive and gated on job_type=='scan'. directed_scan and existing job types are untouched. Wiring rotor_scan_1d through the pipe is a follow-up.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 64.58%. Comparing base (a08d314) to head (aad3328).
⚠️ Report is 1 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1018      +/-   ##
==========================================
+ Coverage   64.55%   64.58%   +0.03%     
==========================================
  Files         119      119              
  Lines       39788    39791       +3     
  Branches    10307    10308       +1     
==========================================
+ Hits        25684    25701      +17     
+ Misses      11123    11107      -16     
- Partials     2981     2983       +2     

Flag	Coverage Δ
functionaltests	64.58% <ø> (+0.03%)	⬆️
unittests	64.58% <ø> (+0.03%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ 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.

[Copilot]

Pull request overview

Adds first-class support for job_type='scan' in ARC’s ASE adapter by running a full 1D relaxed torsional scan within a single ASE script process and emitting energies/angles into output.yml for ARC’s existing YAML scan parser.

Changes:

• Implements relaxed 1D torsion scan machinery in ase_script.py (rotating-top detection, bidirectional hysteretic walks, merge + diagnostics, Hartree output).
• Wires ASEAdapter.write_input_file() to serialize torsions and scan_res for scan jobs.
• Adds hermetic unit tests (EMT-based) covering rotating-top resolution, scan grid/endpoint behavior, parser round-trip, and guardrails.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 3 comments.

File	Description
arc/job/adapters/scripts/ase_script.py	Adds the in-process relaxed torsion scan implementation and hooks it into job_type == 'scan' execution/output shaping.
arc/job/adapters/ase_adapter.py	Ensures scan-specific inputs (torsions, scan_res) are written to input.yml for ASE script execution.
arc/job/adapters/ase_test.py	Adds unit tests validating scan input writing, rotating group detection, scan execution with EMT, and YAML parser round-tripping.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.