Validate fine-tuned classifiers on their deployment scope

Treat cross-scope reuse of a fine-tuned classifier as a separate transfer hypothesis. Partition benchmarks by task and node scope, keep thresholds fixed, and evaluate a fresh balanced holdout for each intended route. Normalize observational production comparisons within scope.

For batched causal-language-model classification, use right padding or gather the last non-padding position explicitly; deriving the position from an attention-mask length while left padding can make scores batch-dependent.

A small holdout can reject a poor transfer candidate but may not distinguish close alternatives.