[mpvginde]

Stacked on #44 (and so on #39–#42). Like those, this PR targets refactor/alignment, so until the stack merges the diff includes its commits. To see only this PR's changes, use 68cea40...feat/ensemble-alignment (5 commits). I'll rebase if the stack changes in review.

Why

Time alignment and interpolation already carry a member dimension through. But every run with an ensemble then crashed in broadcast_nans, which the pipeline runs by default. This PR makes the alignment step work for ensembles, on their own or mixed with deterministic forecasts and observations. Verification (chunking of member in Metric, and which metrics apply to which uncertainty type) comes in a follow-up PR.

Changes

• Uncertainty predicates: ds.mx.is_deterministic(), is_ensemble() and is_quantile(). Like the space and time predicates, they read the trait on every call.
• broadcast_nans no longer crashes when an ensemble meets a dataset without member. The uncertainty dimension, taken from the trait, is never matched between datasets. If any member is missing at a point, the ensemble counts as missing there, and the mask goes to all its members and to the other datasets. This keeps the ensemble size constant over the verification sample, and ensembles of different sizes work together. Any other difference in dimensions on a shared variable raises a ValueError. Please review the decision in ADR-003 (status: proposed). Its main cost is that one missing member discards the whole ensemble at that point.
• Delaunay no longer silently drops variables whose last dimension isn't grid_index, such as an ensemble with member after grid_index. point_index takes the place of grid_index, and the other dimensions keep their order.
• Tests: ds_fcst_ens and ds_grid_ens fixtures, whose member coordinate has standard_name="realization" and units="1". Tests cover time alignment with ensembles (pairwise in all directions, and global), both interpolators with member first or last (numpy and dask), a new tests/test_nans.py, and a mixed deterministic + ensemble Runner.align.

Breaking: broadcast_nans now needs mlwp_uncertainty_trait on every dataset. Datasets from mlwp-data-loaders always have it.

Testing

• 134 tests pass (100 on chore: type annotations and numpy docstrings everywhere, enforced in lint; add AGENTS.md #44), and each commit passes on its own. 12 of the new tests fail on the code before this PR.
• All pre-commit hooks pass (ruff, numpydoc, codespell).
• Only tested on synthetic data so far, not yet on a real ensemble run.

Not in this PR

• Verification of ensembles: xskillscore.crps_ensemble still fails because Metric splits member into chunks, and every metric is still applied to every model.
• Transformations from ensemble to deterministic (ensemble mean, member selection).
• Conventions for the member coordinate in the specs: Uncertainty trait: quantile is not a CF standard name, and the ensemble member coordinate is underspecified mlwp-data-specs#34 and Uncertainty trait: forbid uncertainty dims in deterministic, enforce member/quantile conventions mlwp-data-specs#35.

🤖 Generated with Claude Code