Fix bi-rad singlet composite jobs silently running a restricted reference

[alongd]

What

is_species_restricted() returned early for REFERENCE_AGNOSTIC_METHOD_TYPES (composite,
force-field, semiempirical) before reading number_of_radicals, so the bi-rad-singlet branch
below it was unreachable for exactly those method types. Move the number_of_radicals > 1 check
ahead of that early return.

Why

An open-shell singlet biradical is declared in ARC as an ARCSpecies with multiplicity=1 and
number_of_radicals=2 — that is what the ARCSpecies docstring prescribes, and it is the only way
to express the state. The existing code already knew what to do with it; its own comment says so:

# don't run unrestricted for composite methods such as CBS-QB3, it'll be done automatically if the
# multiplicity is greater than one, but do specify uCBS-QB3 for example for bi-rad singlets.

That comment describes a case the function could never reach. Composite methods take
method_type == 'composite', so G4/CBS-QB3 returned True (restricted) at the top, and the
branch the comment documents never ran. A declared bi-rad singlet therefore went out as plain g4
with a restricted reference — no u prefix — and converged to the wrong electronic state
without warning. Nothing errors; the thermochemistry is simply wrong.

The asymmetry is the whole point: for a composite method an unrestricted reference is applied
automatically when multiplicity > 1, but a bi-rad singlet's multiplicity is 1, so the method
gets no signal and ARC has to say uG4 explicitly.

What changed

arc/job/adapters/common.py — number_of_radicals > 1 is now evaluated first and returns
unrestricted, then the reference-agnostic early return, then the multiplicity > 1 case. The
existing # don't run unrestricted for composite methods... comment is replaced by one that
explains why the order matters, so the next reader does not "simplify" it back.

Unaffected: plain closed-shell singlets (number_of_radicals unset or 1) and composite jobs at
multiplicity > 1 both take exactly the paths they did before.

Testing

arc/job/adapters/common_test.py — 61 passed, with new cases covering a composite bi-rad singlet
(unrestricted), a composite closed-shell singlet (restricted), and a composite triplet (unchanged).

Verified end to end on the route section: a g4 job for a species with multiplicity: 1, number_of_radicals: 2 now writes ug4.

[Copilot]

🟡 Changes recommended

Preserve force-field reference-agnostic handling before approval.

Get a fresh assessment by requesting another Copilot review.

Pull request overview

Fixes unrestricted-reference selection for composite bi-radical singlet jobs.

Changes:

• Prioritizes number_of_radicals > 1.
• Adds regression coverage for composite and Gaussian route generation.

File summaries

File	Summary
arc/job/adapters/common.py	Reorders reference selection; a critical finding notes this can produce uuff for force-field jobs.
arc/job/adapters/common_test.py	Adds composite biradical, closed-shell, triplet, and route-generation tests.

Review details

• Files reviewed: 2/2 changed files
• Comments generated: 1
• Review effort level: Lite

---

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

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 66.18%. Comparing base (575d216) to head (51f93ce).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1055      +/-   ##
==========================================
- Coverage   66.26%   66.18%   -0.09%     
==========================================
  Files         122      122              
  Lines       41824    41827       +3     
  Branches    10751    10751              
==========================================
- Hits        27716    27682      -34     
- Misses      11062    11090      +28     
- Partials     3046     3055       +9     

Flag	Coverage Δ
functionaltests	66.18% <ø> (-0.09%)	⬇️
unittests	66.18% <ø> (-0.09%)	⬇️

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.

[alongd]

Added a second, related fix to this branch (51f93ce), since it is the same failure — a job that runs the wrong electronic state and reports it as a normal result — reached by a different route.

ARCSpecies.__init__ and ARCSpecies.from_dict disagree on the multiplicity of the same species:

xyz = "<the optimized geometry of C[C]C#N>"
ARCSpecies(label="s", smiles="C[C]C#N", xyz=xyz).multiplicity                      # 3
ARCSpecies(species_dict={"label": "s", "smiles": "C[C]C#N", "xyz": xyz}).multiplicity  # 1

__init__ reads multiplicity/charge off the molecule built from the SMILES/adjlist/InChI (species.py:507-508) and only then may overwrite .mol from the coordinates, guarded by regen_mol and not keep_mol. from_dict runs mol_from_xyz() first with no guard (1011-1014) and reads the multiplicity afterwards (1024-1025). The two differ by statement order alone. Perception picks an electronic state, and for an open-shell species it routinely picks a different one: C[C]C#N perceives from its own optimized coordinates as C[C][C][N].

A deck travels the from_dict path, so this is not hypothetical — it silently ran 13 of 31 open-shell species in a controlled A/B study at the wrong multiplicity. The jobs converged and terminated normally.

The fix captures multiplicity and charge from the structure-derived molecule before the override, falling back to the perceived molecule only when no structure was given at all. An explicitly declared multiplicity still wins. arc/species/ + arc/main_test.py + arc/job/adapters/common_test.py run serially: 446 passed, 1 failure (main_test.py::TestARC::test_as_dict) that is present on this branch before the change as well.