add Winkler metric

[ContiPaolo]

Closes #286 by implementing Winkler metric

Summary of Changes

• Added Winkler Interval Score (WinklerScore) as a new ensemble metric in ensemble.py.
• Registered the metric in the metric registry and evaluation scripts.
• Added unit tests for correctness and parameter validation.
• Documented the metric with explicit shape conventions and mathematical definition.

Winkler Interval Score Definition

The Winkler interval score evaluates the sharpness and coverage of central prediction intervals. For significance level $\alpha \in (0, 1)$, the score for a single point is:

$$W_\alpha = (u - l) + \frac{2}{\alpha}(l - y)\mathbf{1}(y < l) + \frac{2}{\alpha}(y - u)\mathbf{1}(y > u)$$

where:

• $l$ is the lower bound of the central $(1-\alpha)$ prediction interval (ensemble quantile at $\alpha/2$)
• $u$ is the upper bound (ensemble quantile at $1-\alpha/2$)
• $y$ is the observed value

Lower values are better: narrow intervals are rewarded, and misses are penalized in proportion to their distance outside the interval.

The implementation supports flexible reduction over spatial/temporal dimensions and returns the mean Winkler score by default.

[Copilot]

Pull request overview

Adds a new ensemble evaluation metric (Winkler Interval Score) to quantify prediction interval sharpness/coverage tradeoffs, integrating it into the metrics registry, evaluation CLI defaults, and unit tests.

Changes:

• Implemented WinklerScore as a new BTSCMMetric ensemble metric based on ensemble quantiles.
• Registered the metric in autocast.metrics exports and the encoder/processor/decoder eval script metric registry/default list.
• Added unit tests covering a hand-computed reference case and alpha parameter validation.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

File	Description
src/autocast/metrics/ensemble.py	Adds WinklerScore implementation and its docstring/shape conventions.
src/autocast/metrics/__init__.py	Exports and registers WinklerScore in ALL_ENSEMBLE_METRICS.
src/autocast/scripts/eval/encoder_processor_decoder.py	Makes winkler available in the eval CLI and adds it to default metrics.
tests/metrics/test_ensemble.py	Adds correctness and input-validation tests for WinklerScore.

---

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

[ContiPaolo]

@sgreenbury Comments on computational cost: Winkler (#286) vs Variogram(#285)

• The Variogram score can become computationally and memory intensive when applied to the full $(B, T, S, C)$ tensor, because it computes all pairwise differences across the selected vector dimensions, leading to $O(D^2)$ scaling (where $D$ is the flattened spatial/temporal/channel size).
• The Winkler score does not suffer from this problem. It only computes two quantiles (lower/upper) along the ensemble axis $M$, then applies elementwise operations to produce a tensor of shape $(B, T, S..., C)$. There are no pairwise operations across spatial/temporal/channel axes.
• The only scenario where Winkler can produce numerically large values is if $\alpha$ is set very small (since penalties scale as $2/\alpha$), but this does not affect memory or computational scaling.

[sgreenbury]

LGTM!