Resolve which holdout metric to compare on. When the policy pins a concrete
metric name, that wins. With 'auto', prefer the default higher-is-better
metric for the incumbent's problem type that is present on BOTH models, then
fall back to the first numeric metric shared by both, then to the first default
name (so the caller still gets a stable key even if values are absent).
Parameters
configured: string
the policy's comparisonMetric ('auto' or a metric name)
Resolve which holdout metric to compare on. When the policy pins a concrete metric name, that wins. With
'auto', prefer the default higher-is-better metric for the incumbent's problem type that is present on BOTH models, then fall back to the first numeric metric shared by both, then to the first default name (so the caller still gets a stable key even if values are absent).