Views of each chunk: as is; NFKD + marks/format/fillers dropped + Cyrillic/Greek lookalikes mapped + dashes normalised, with whitespace collapsed (acme corp, \s) or removed (acme-\nclient); and a soft hyphen read as a dash. Chunks of 1 MiB with a 4 KiB overlap bound the folding's allocation (it was ~14x the input: 919 MiB for a 64 MiB note). Documented limits: leetspeak, punctuation splits, encodings and homoglyphs outside a small table. Mutation-checked. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>