[leifdenby]

Addresses bullet 2 of my notes on #30: "the config.py in utils implies a config schema, should we make one explicit?"

Note

This builds on #41. Both PRs come from my fork, and GitHub can't stack PRs across forks, so this one targets refactor/alignment and also shows #41's commits. Only the last two commits are new here. Once the others are merged, the diff will show only this PR's changes.

Changes

• mxalign/pipeline/config.py: the untyped Config dict wrapper is replaced by pydantic models, PipelineConfig plus one model per section (DatasetConfig, TransformationConfig, AlignmentConfig, TimeAlignmentConfig, SpaceAlignmentConfig, AlignmentSaveConfig, VerificationConfig, MetricConfig, SaveConfig, DatesConfig). Load with load_config(path_or_dict).
  • Strict sections reject unknown keys. Examples: a typo'd top-level section, alignment.time.return_as, or alignment.save.datasets: some.
  • Sections whose extra keys are forwarded to a function (loader, transformation, metric, save) accept them, and they're available as .kwargs.
  • dates → files expansion now happens in the validator, as before. A per-dataset dates overrides individual fields of the top-level one, and incomplete dates give a clear error. files may also be a list of patterns.
• Runner reads typed attributes instead of .pop()/.get() on dicts. This also fixes align() mutating the config: it popped reference, so calling it twice failed. verify() is skipped when there's no verification section.
• Adds the pydantic>=2 dependency. The uv.lock change only adds pydantic and its three dependencies.
• New tests/test_config.py: examples/config.yaml validates, plus tests for unknown and missing keys, pass-through kwargs, and dates override and validation. The existing runner/dates tests are updated. The "unknown alignment.time option" check now comes from the schema.
• README, CHANGELOG and examples/advanced.ipynb updated: runner.config is now the model, not a callable.

Behaviour changes worth checking

• A config without alignment is now rejected at load time. The runner already required it.
• inputs values of a metric must be forecast or reference. Previously any value other than reference was silently treated as forecast.
• alignment.space: {} now runs space alignment with default options. Previously an empty dict skipped it.

100 tests pass.

🤖 Generated with Claude Code