// Minimal Swift contract for the OpenMed Persian PII CoreML INT4 export.
// Tokenization is intentionally external: use the bundled tokenizer files or a
// compatible WordPiece tokenizer, then feed one padded 256-token window at a time.
//
// Required input tensors:
// - input_ids: Int32 shaped [1, 256]
// - attention_mask: Int32 shaped [1, 256], filled with 1 for every position
// - token_type_ids: Int32 shaped [1, 256], usually all 0
//
// Required postprocessing:
// - ignore special tokens and zero-length/pad offsets
// - trim whitespace from predicted spans
// - merge overlapping sliding-window spans
// - add regex/rule assists for phone, email, national ID, postal code, card, date