The scorer is now script-aware. Japanese prompts were previously abstained
(unscored) by the non-English guard; they are now scored on a dedicated path,
dictionary-free and fully offline — no morphological-analyser/MeCab dependency,
consistent with the module contract.

Design: detect_lang() routes on kana presence. Japanese is tokenized by
character-class run (tokenize_ja, no spaces to split on) for diversity/
repetition, sized by a char/2 word-equivalent, and matched against parallel
JA_* slot lexicons by substring (a space-less script has no word boundaries).
The slot rubric, multiplicative core, guards, and aggregation are unchanged —
they consume feature counts, so they work for either script. Readability
indices are English-specific, so Japanese uses a sentence-length clarity proxy
(ja_clarity); the English Flesch/FK/Fog path is untouched. Sentence counting
now also splits on 。！？. Languages other than English/Japanese still abstain.

Japanese calibration parallels English closely (vague~12 nascent, tactical~53,
rich~56, loaded~58 proficient). English calibration is byte-identical — the JA
path is fully isolated. PromptScore gains a public  field so callers can
label the score. Known limitation (documented): substring matching lets a
precise action stem (変更) also match inside a prohibition (変更しない), so a
pure-prohibition JP prompt scores a little objective credit it wouldn't in
English. 49 module + 5 integration tests green; clippy clean both configs.