[leifdenby]

Every function now has type annotations on all arguments and return values and a numpy-style docstring. A lint check keeps it that way, and AGENTS.md describes the project's coding conventions and development tools.

Note

This builds on #41 and #42, a stack of follow-ups to #30 (refactor/alignment). All of these come from my fork, and GitHub can't stack PRs across forks, so this one targets refactor/alignment (#30) and also shows their commits. Only the last commit is new here.

Enforcement (runs in pre-commit, and so in the existing lint CI)

• pyproject.toml enables ruff's ANN rules (annotations) and D rules (docstrings, numpy convention), with no exemptions for tests.
• The new numpydoc-validation hook (v1.11.0) checks that every parameter, return value and so on is documented. It runs all checks except EX01/SA01/ES01 (Examples, See Also and Extended Summary are optional). __init__ is excluded because the class docstring documents its parameters.
• src/mxalign/loaders/ifs_forecast.py is excluded from both: it's orphaned and imports modules that no longer exist, and it should be removed or ported to mlwp-data-loaders separately.

Code (src/, tests/, examples/myloader.py)

• Every module, class, function, method and test now has full annotations, with no Any; pass-through **kwargs are typed object. Each also has a numpy docstring (Parameters / Returns / Raises).
• pydantic config models document their fields under Attributes, because numpydoc can't see pydantic's generated constructor.
• No behaviour change. An AST comparison with docstrings, annotations and imports stripped shows only three code differences:
  • mxalign.__all__ is now sorted.
  • dict(...) became {...} in two tests.
• All 100 tests pass.

AGENTS.md: package layout, development tools (uv, pytest, pre-commit), coding conventions (annotations, numpy docstrings with an example, no mutation of inputs, registries, config schema), tests, CHANGELOG and ADRs.

Possible bugs found while documenting, not fixed here (each could be its own issue):

• Runner.verify: if verification.save is set but there are no metrics, self.metrics is never assigned, so saving raises AttributeError.
• Runner.load_datasets: if none of a dataset's files exist, an empty list is passed to the loader.
• DatasetPath: for a dataset that is neither forecast nor observation, year/month/day are never set, so substitute raises AttributeError.
• MxAccessor.align_time_with: can pass a list/timedelta lead_time to align_forecast_to_forecast, which only accepts "reference"/"intersection"/"union".
• utils/traits.update_*_trait: the validation report is discarded, so invalid traits are never reported.
• broadcast_nans: an empty dict input comes back as an empty list.
• The transformation name kelvin_to_celcius is misspelled (should be "celsius"), and it's public.

🤖 Generated with Claude Code