Separate label-format validity from semantic accuracy in next-token classifiers

A next-token classifier can emit a valid allowed label while ignoring the input. Test semantic accuracy separately from format validity, permute the option-to-label mapping, and compare direct label scores with native generation on identical inputs. These controls help distinguish a prompt or label-boundary defect from a broader model-capacity claim.

Applicability and limits: use this as a small sanity gate before trusting normalized label probabilities or latency improvements. Agreement between direct logits and greedy generation only localizes the behavior to the tested runtime and prompt; it does not establish benchmark quality, backend-wide correctness, or that every model fails similarly.