Morpheus Tokenizer Uses Neural Morpheme Boundaries for Turkish NLP

Morpheus models Turkish morpheme boundaries using a differentiable Poisson-binomial dynamic program over per-character boundary probabilities, producing lossless tokenization where decode(encode(w)) = w holds by construction. The same forward pass generates both segmentation and word embeddings, avoiding the corpus-statistics fragmentation of BPE/WordPiece on agglutinative languages.