All-forms underline (mark one, light up every inflection)

Parent: key-designs

The distinctive reading feature: save one inflection of a word, and every inflected form of its lemma is underlined throughout the book. (Internally "Task #2".)

Form-sets. services/formMatching.ts builds a lemma's match set: entryFormSet = the LLM/dict-harvested forms ∪ the surface word ∪ the lemma; normalizeForms lowercases/trims and drops multi-token surfaces (so "le château" can't match a bare page token, but "château" still does; "ils/elles méritent" can't spuriously match). buildMarkedWordsMap (services/displayUtils.ts) then keys a map by every form, and the reader underlines a page token when markedWords.get(token.toLowerCase()) hits (ClickableText.tsx).

The tokenizer has to agree with everything else. services/tokenizationService.ts is a Unicode-aware splitter (Latin punctuation; CJK / Hiragana / Katakana / Hangul each their own class, per-char for CJK) with a French elision merge pass that re-joins l', d', j', qu' + the following word into one token — matching the tokenization rules baked into the LLM prompt. This matters because token boundaries must be identical across three consumers: the underline matcher, POS coloring, and the LLM word-analysis. A disagreement would underline the wrong span or mis-color.

German separable verbs get special handling here too (services/separableVerbService.ts heuristic on the client), and properly in the Python engine (linguistic-two-engines).

Why it's load-bearing, not cosmetic: the underline is the feedback surface of the whole vocabulary-loop — it's what makes saved words visible while reading, and it depends on form-sets + a tokenizer that agrees with the linguistic layer. Get the tokenizer wrong and the core feature silently mis-paints.