Optional analyzer-level NLP engine for Bumblebee token-classification serving.
The engine attaches normalized model outputs to Obscura.NLP.Artifacts so
recognizers can consume model evidence without each recognizer invoking model
serving independently.