NaN-handling for Energy Score

[frazane]

Brings nan_policy support to the multivariate Energy Score family, following the same approach introduced for the ensemble CRPS in #118. This is the first multivariate metric to gain NaN handling; the pattern is meant to carry over to the remaining multivariate scores.

As with CRPS, a member is treated as invalid if any of its variables (or its weight) is NaN, and nan_policy accepts propagate (default), raise, and omit. The omit policy zero-weights invalid members and routes the score through the existing weighted estimators, so it reduces to wiring rather than new numerics. NaN in the observations always propagates.

What changed

• Added nan_policy to es_ensemble, twes_ensemble, owes_ensemble and vres_ensemble.
• Reworked the apply_nan_policy_ens_mv helper (added in NaN-handling for ensemble CRPS #118 but unused and untested) to mirror the univariate apply_nan_policy_ens_uv: it now folds invalid members into the ensemble weights and returns the aligned ens_w.
• Fixed two pre-existing bugs in the weighted numba energy kernels, both only reachable once ens_w is actually exercised on the numba backend (which the omit path now does): a scalar-indexing error (ow = ow[0]) in the outcome-weighted and vertically-rescaled kernels, and a wrong normalisation (np.mean(fw) instead of np.sum(fw * ens_w)) in the outcome-weighted kernel.
• Added tests covering all three policies, including per-row equivalence of omit against the score computed over the surviving members (for both the nrg and fair estimators), NaN weights, and the weighted-ens_w paths that the kernel fixes depend on.

Notes

• omit is not implemented for the akr and akr_circperm estimators (they are roll/permutation-based and cannot be expressed through member weighting), matching how CRPS treats its akr/int estimators. They raise NotImplementedError.
• The same ow = ow[0] pattern still exists in the weighted multivariate kernel-score gufuncs (core/kernels/_gufuncs_w.py); left as a follow-up since it is outside the Energy Score scope here.

[Copilot]

Pull request overview

This PR adds nan_policy handling to the multivariate Energy Score family (ES/twES/owES/vrES), mirroring the existing CRPS NaN-handling approach. It wires NaN member omission through the existing weighted estimators (by zero-weighting invalid members) and fixes two weighted numba energy kernels that become exercised by the new omit path.

Changes:

• Added a nan_policy parameter to es_ensemble, twes_ensemble, owes_ensemble, and vres_ensemble, and integrated it into the scoring flow (including proper interaction with ens_w).
• Reworked apply_nan_policy_ens_mv to align/mask ens_w and to support propagate / raise / omit consistently with the univariate helper.
• Fixed weighted numba energy gufunc issues (scalar indexing of ow, and wbar normalisation for outcome-weighted energy) and added comprehensive tests across policies and weight scenarios.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
tests/test_wenergy.py	Adds NaN-policy tests for owES/vrES/twES and a uniform-ens_w equivalence guard for the weighted kernels.
tests/test_energy.py	Adds NaN-policy coverage for es_ensemble, including NaNs in forecasts, NaNs in ens_w, and estimator-specific omit behavior.
scoringrules/core/utils.py	Reworks apply_nan_policy_ens_mv to return aligned ens_w and implement omit-by-zero-weighting for invalid members.
scoringrules/core/energy/_gufuncs_w.py	Fixes weighted numba ow/vr energy gufunc bugs (scalar indexing and wbar computation).
scoringrules/_energy.py	Threads nan_policy through public Energy Score APIs and routes omit cases through the weighted code paths.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.