The pipeline that produces fluent nonsense

A pipeline caches tokenized text to avoid re-tokenizing a large corpus. Stage one tokenizes with a tokenizer loaded from a local path and writes integer arrays to Parquet. Stage two loads those arrays and runs them through a model.

After someone upgraded the model checkpoint to a newer release from the same family, outputs became fluent, grammatical, and semantically unrelated to the input. No error is raised. Sequence lengths are unchanged. Token id ranges look normal.

Perplexity computed on the cached ids against the new model is roughly 900.

Diagnose this from the evidence given. Then tell me why nothing in the system raised an error, and what invariant the pipeline should have been asserting all along.