Add L1 (CRPS-consistent) ssr_bias_l1 metric to one-step ensemble aggregator

[mcgibbon]

Adds an L1 (CRPS-consistent) analog of the existing L2 ssr_bias to the ensemble aggregator, so 1-step ensemble under-dispersion can be read in units that match the fair-CRPS / energy-score training loss rather than MSE. It is exposed on both the one-step ensemble aggregator and the train/val aggregator path.

The metric is ssr_bias_l1 = spread_L1 / skill_L1 - 1 (target 0 at calibration, same convention as the L2 metric), where the two terms are the two CRPS terms:

• spread_L1 is the fair mean pairwise absolute difference, 1/(N(N-1)) sum_{i!=j} |x_i - x_j| (the i!=j correction plays the role unbiased variance plays for the L2 spread).
• skill_L1 is the member MAE, mean_i |x_i - y| (member-vs-obs, not ensemble-mean-vs-obs). At calibration E|X - X'| = E|X - y|, so the ratio approaches 1. Unlike the L2 skill's mse - var/N, the member MAE is already unbiased and gets no finite-N correction: MAE-of-the-ensemble-mean has no closed /N decomposition, so there is no L2-style analog to subtract.

The numerator and denominator sums accumulate across batches and the ratio is formed once in get(), so it is not a per-batch-averaged ratio. Prescribed cells (every member equals the target: zero spread and zero error) report 0 rather than the -1 floor, matching SSRBiasMetric.

Off by default: enable_ssr_bias_l1 gates building the family, so existing runs' logged keys are unchanged.

Changes:

• fme.ace.aggregator.one_step.ensemble.SSRBiasL1Metric: new ReducedMetric accumulating the spread and skill sums and forming the ratio in get(), with the prescribed-cell 0/0 handling carried over from SSRBiasMetric.

• _EnsembleAggregator: enable_ssr_bias_l1 flag builds the ssr_bias_l1 family (registered in _get_variable_metrics, added to _diverging_metrics), emitting ssr_bias_l1/<var>, ssr_bias_l1/mean_map/<var>, and (for target="norm") ssr_bias_l1/channel_mean via the existing area-weighted-mean and mean-map plumbing.

• get_one_step_ensemble_aggregator, EnsembleMetricConfig, OneStepEnsembleMetricConfig: gain enable_ssr_bias_l1: bool = False.

• fme.ace.aggregator.train.TrainAggregatorConfig: gains enable_ssr_bias_l1: bool = False, forwarded into the _EnsembleAggregator built when ensemble_metrics is enabled, so the L1 metric is available on the train/val aggregator path (e.g. val/ensemble/ssr_bias_l1).

• Tests added

• If dependencies changed, "deps only" image rebuilt and "latest_deps_only_image.txt" file updated

[mcgibbon]

Claude: Independent pre-review (fresh Opus agent, no authoring context; worktree/env-slot isolated on the PR head bbe493f). Verdict: ready for review — no blocking issues found, no fixes applied.

What I checked:

1. L1 math is correct. spread_L1 is the fair pairwise mean: (gen.unsqueeze(2) - gen.unsqueeze(1)).abs().sum(dim=(1,2)) / (N*(N-1)) — the i==j terms are exactly 0 so this equals the i!=j sum, giving the unbiased E|X-X'|. skill_L1 is member-vs-obs MAE (gen - target).abs().mean(dim=(0,1,2)), not ensemble-mean, and correctly gets no -var/N correction. Independent numeric check: reference spread=5.333, skill=3.333, bias=0.6 matched the metric to float precision (0.6000001), and a 20k-sample iid calibrated ensemble gave bias -0.0017 (→0). ✓

2. Ratio-at-end. Numerator/denominator accumulate in _total_spread/_total_skill; ratio formed once in get(). The test_..._forms_ratio_at_end_not_per_batch test genuinely distinguishes the two orderings: it asserts == (2+10)/(4+5)-1 (0.333) and != 0.25 (the per-batch-ratio mean). A per-batch implementation would fail it. ✓

3. Cross-batch weighting matches SSRBiasMetric: per-batch means summed across batches, ratio at end, so n_batches cancels and each batch is equally weighted. Consistent and defensible. ✓

4. Default-off invariance. The family is registered in _variable_metrics only when enable_ssr_bias_l1=True; _get_data iterates families generically, so nothing leaks when disabled. The unconditional _diverging_metrics membership is inert with no family present. Test confirms no ssr_bias_l1 keys in logs/dataset and existing ssr_bias/a unchanged. ✓

5. Prescribed-cell 0/0 reports 0, matching SSRBiasMetric. Correctly uses linear MAE units for the floor (_PRESCRIBED_MSE_RTOL * skill.max(), not .square().max()). Test covers prescribed→0 and zero-spread-with-error→-1. ✓

6. Plumbing/perf. enable_ssr_bias_l1 threaded through both EnsembleMetricConfig and OneStepEnsembleMetricConfig, both .build() paths, and get_one_step_ensemble_aggregator → _EnsembleAggregator; mean_map/channel_mean flow through the generic loop. Pairwise materializes (batch, N, N, time, lat, lon) — O(N²), fine for the small one-step ensemble sizes and consistent with existing style.

Verification run in the env slot: test_ensemble.py 33 passed; test_main.py + test_train.py 32 passed (no regressions); pre-commit (ruff, ruff-format, mypy) all pass on both changed files.

Two cosmetic non-issues (mentioned only for completeness, no change needed): SSRBiasL1Metric._n_batches is tracked but unused in get() — but SSRBiasMetric does the same; and the shared _PRESCRIBED_MSE_RTOL constant name carries "MSE" while reused as a generic rtol for the L1 floor.

[mcgibbon]

I'll make use of this in an experiment before we consider reviewing/merging.