normalize.zig: preserve Han + Hiragana, fold katakana->hiragana + full-width ASCII->ASCII (NFKC-lite); Latin fold table unchanged; VERSION -> v2-cjk-bigram (forces re-index via the snapshot loader's normalize_version gate). prefixes.zig: one shared UTF-8-codepoint-safe tokenizer (emitTokenKeys) used by BOTH builder (displayPrefixKeys) and serving (queryKeys). Segments Latin vs CJK runs; CJK runs emit adjacent bigrams + a leading unigram (no full unigram indexing); Latin runs unchanged. 100/100 index tests pass. Refs docs/i18n-cjk-search.md. Not yet: scoring boosts, manifest cjk_key_count + promote floor gate, full re-index (separate go/no-go). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>