Split ASE/thermo.py into a package, one module per concern

[isayev]

What

1895 lines became four modules under Auto3D/ASE/thermo/:

module	contents
properties	molecular inspection — geometry class, symmetry number, multiplicity, display name. No model, no calculator.
calculator	the ASE calculator fronting an Auto3D model, and the conversions into it
vibrations	Hessian, Eckart/Sayvetz projection, frequencies
driver	the per-record sequence and the run over a file

Dependency order is properties, calculator ← vibrations ← driver, read off the import graph rather than asserted. Each of the 33 top-level statements was compared by ast.dump against its original: byte-identical, none lost, none invented.

A package, not four sibling modules — and not for tidiness

get_logger(__name__) now yields Auto3D.ASE.thermo.properties and friends, which are children of Auto3D.ASE.thermo. Records propagate, so the fifteen-odd tests capturing on logger="Auto3D.ASE.thermo" keep seeing warnings raised anywhere in the four.

Siblings named thermo_properties would have broken every one of them — and broken them silently, because a log-capture assertion that stops seeing records fails by never firing. Demonstrated directly rather than argued: the child propagates, the sibling does not.

__init__.py re-exports calc_thermo alone, permitted by the package-barrel rule for exactly the reason isomers already records: api.rst documents Auto3D.ASE.thermo.calc_thermo at that package path.

The bug this split creates, and the test that didn't exist

_symmetry_default_warned de-duplicates _symmetry_number's sigma=1 warning to once per run, and calc_thermo clears it at the top of each run. The flag lives with _symmetry_number in properties now — so global _symmetry_default_warned in driver would bind a name in driver that _symmetry_number never reads. The reset silently stops working and the warning fires once per process instead of once per run. The assignment goes through the module object instead.

Mutation-testing is what found the real gap. The naive global version passed the entire suite: every existing test of this warning sets the flag itself, so they pass either way, and the per-run reset was never covered at all. test_calc_thermo_resets_the_symmetry_warn_once_flag_in_properties is the missing guard, and the mutation now fails it with a message naming the consequence.

One more consequence worth knowing

create_model is now imported into two modules, so calc_thermo reaches it twice — via _load_hessian_model (driver) and via model_name2model_calculator (calculator). A single monkeypatch used to cover both. One test measuring the ASE relaxation was silently loading a real model until both bindings were patched.

The test repointing was mechanical but not blind

Every misrouted monkeypatch located a name exactly where this split says it belongs — Atoms and create_model with the calculator that constructs them, VibrationsData and _ASE_HAS_VIB_SELECTION with the Hessian code, _electron_count and the warn-once flag with the molecular properties. The tests were an independent check on the grouping, and they agreed with it.

Docs

docs/source/migration-3.0.rst names projected_vibrations by dotted path in a live description of current behavior; updated to its new module. It is the only doc reference to a moved name — api.rst points at calc_thermo, which did not move.

Verification

• 1771 passed, 1 skipped, 74 deselected (+1 new guard), randomized order.
• mypy: 68 errors, unchanged. 23 files instead of 22 and 79 checked instead of 75, because one file became four plus an __init__.
• All 33 moved statements AST-identical to their originals.
• Mutation-tested both hazards: the naive global reset fails the new test, and the logger-hierarchy property the package decision rests on is demonstrated rather than assumed.