[yyexela]

Adds a default-on inference metric that measures slow drift of the global-mean bias away from the target, one slope per sample (initial condition or ensemble member) per variable. The per-sample slopes are exposed via InferenceEvaluatorAggregator.get_global_mean_drift_slopes() so a downstream trainer can test whether the drift is significant using the spread across initial conditions. That accessor has no in-repo caller yet; it is a hook for a downstream trainer.

The aggregator accumulates running OLS sums (n, Σt, Σt², Σy, Σty) in O(n_samples × n_vars) memory and tracks the recorded step count as a scalar (identical on every rank). Slopes are NaN when n_time * timestep < 730 days, matching the annual-metric minimum. Samples are gathered across data-parallel ranks at finalize time.

record_batch(data)                        # per sample s, per variable v
  y[s, t] = area_weighted_mean(gen - target)[v]
  t       = years since sample's first recorded step
  accumulate n, Σt, Σt², Σy, Σt·y

finalize
  data_parallel_gather(stats)
  slope[s, v] = (nΣty − ΣtΣy) / (nΣt² − (Σt)²)
  slope = NaN where n_time * timestep < 730 days

Logs (under the metric name global_mean_drift):

• global_mean_drift/slope/{v}: sample-mean slope (units of v per year)
• global_mean_drift/t/{v}: mean / (std(ddof=1) / sqrt(n_samples)); NaN when n_samples < 2
• global_mean_drift/max_abs_t: max over variables of |t|, ignoring NaN

global_mean_drift_diagnostics.nc holds slope(sample, variable) and slope_units(variable).

The metric is added only to InferenceEvaluatorAggregatorConfig — it is not added to LegacyFlagInferenceEvaluatorAggregatorConfig or the coupled InferenceEvaluatorAggregatorConfig (fme/coupled/aggregator.py).

Changes:

• fme.ace.aggregator.inference.global_mean_drift: new GlobalMeanDriftAggregator and GlobalMeanDriftConfig (default-on, skipped for rollouts ≤ 730 days). GlobalMeanDriftConfig is exported from fme.ace.

• fme.ace.aggregator.inference.data: new PerSampleSlopes protocol and MetricBuildResult.per_sample_slopes slot, so the build loop hands the aggregator to the top level without isinstance checks.

• fme.ace.aggregator.inference.main: registers the config in the MetricConfig union, adds global_mean_drift field to InferenceEvaluatorAggregatorConfig, and adds InferenceEvaluatorAggregator.get_global_mean_drift_slopes().

• fme.core.testing.MockDistributed: adds data_parallel_gather.

• Tests added

• If dependencies changed, "deps only" image rebuilt and "latest_deps_only_image.txt" file updated (no dependency changes)