36. A leading proclitic is not a noun, and it is not a rhyme word either

Context

s_plus_7 and n_plus_7 walk a source text word by word, looking each one up in pack.noun_index and comparing the candidate against the displacement the checker computes. Measured on the branch point, this inverted the verdict rather than merely missing it:

Call — offset: 3, ambiguous_nouns: "strict"	Result	Should be
l'îlot from source l'île — the correct S+7	satisfied: false, changed_a_non_noun	satisfied
l'île from source l'île — the source retyped unchanged	satisfied: true, score 1.0	flagged; strict exists for exactly this

Bare île worked (île → îlot under the same offset), so one word gave two verdicts depending on a preceding l'. d', qu' and the typographic ’ behaved the same. Both rows share the defect: s_plus_7 delegates entirely to n_plus_7's displace and displacement_report.

The cause is noun_index's own lookup, not the tokeniser. word_spans keeps an apostrophe-bearing token whole on purpose — 94 real French words carry an internal apostrophe of their own (aujourd'hui, prud'homme), and splitting every apostrophe unconditionally would misread those. denckring_fr_data's syllable and phoneme lookups (_table_entry) already learned this: try the token whole against the table first, and only on a miss fall back past the last apostrophe. noun_index never got that fallback, so noun_index("l'île") returned None even though noun_index("île") succeeds (index 44740). This is chapter 6's trap 1 again — a lookup normalising differently from the way its table is keyed.

Unlike the syllable table, no noun carries an internal apostrophe — checked against the shipped 44,746-entry list, zero matches. So for this row specifically, an apostrophe in a word handed to noun_index is always an elision boundary, never part of the noun, and the try-whole-first step the syllable path needs is not needed here: splitting is always safe.

Decision

n_plus_7.py gains split_elision(word) -> (prefix, tail), splitting at the last of ' or ’. displace and displacement_report both work on the tail for every lookup, comparison and displacement, and treat the prefix as a separate, inert unit that must match exactly between source and candidate — a new changed_proclitic violation, distinct from changed_a_non_noun, because the fault is the prefix, not the noun.

The prefix is reattached exactly as written, never re-elided against the displaced noun's own initial sound. l'île displaces to l'îlot correctly because both île and îlot start with a vowel. A displacement landing on a consonant-initial noun keeps the same written proclitic — l'aïoli displaced by one entry becomes l'b, which no French speaker would write by hand. Choosing correctly between l'/le/la (and d'/de, qu'/que) needs the displaced noun's grammatical gender and its initial sound class (vowel, mute h, aspirated h), none of which the noun list carries. Building that is its own decision, not folded into this one; tests/test_elision_seam.py pins the literal behaviour so it is a known limitation rather than a silent one.

Language-blind by construction: no English or German word carries an apostrophe, so split_elision is a no-op there and the fix costs nothing outside French.

split_elision lives in core/text.py, not n_plus_7.py, because a second caller needed it the same day. identical_rhyme (core/prosody.py) compares each rhyme scheme line's final word against another's to decide whether they are the same word — l'amour and amour are, and the comparison read them as different, so a minimal pair differing only by a leading l' matched by does_rhyme and by neither identical_rhyme nor does_not_rhyme, scoring 1.0 where the bare pair correctly failed. hemeling/limerick/rhyme_scheme all share core/prosody.scheme_violations, so one fix reaches all three. Sharper than a generic gap: hemeling publishes allow_identical as "Permit a word to rhyme with itself, as French rime riche does" — the row explicitly models French self-rhyme, and the dial was bypassed by the commonest orthographic fact in the language. This caller does not need the try-whole-first step either, for a different reason than noun_index's: it only ever compares one split segment against another split the same way, so even a genuine apostrophe-word compared with itself still comes out correctly identical.

Also fixed the same day: kangaroo_word's own lookup, a different root cause under the same MCP sweep finding. synonym: "école" failed not_a_word under default fold_diacritics, because is_word is not fold-aware — French keeps its accents on purpose (ADR 0009) — and kangaroo_word folded the synonym before asking it, so a diacritic-folded ecole could never match the table's école. Not an elision bug and split_elision is not involved: every other is_word caller in this codebase (paragram, word_ladder, semordnilap) already checks membership on the word as written and folds only for its own scattering/order comparison, so the fix is is_word on params.synonym rather than on the pre-folded copy — bringing the one row that disagreed into line with the rest, not a new decision.

Consequences

The generator inherits the same literal-reattachment limit as the checker. apply("s_plus_7", "l'aïoli", offset=1) returns l'b, and its own checker accepts it — consistent, not obviously grammatical. This is the same shape of cost this project has taken before: ambiguous_nouns's default, monoconsonantal's French y, and it is recorded rather than hidden.

changed_proclitic is a new rule name. Any caller pattern-matching on displacement_report's violation rules for these two procedures gains one more value.

Both related sub-findings from the same MCP sweep finding are fixed alongside this one, recorded above rather than as open items: kangaroo_word's accent-folding bug and identical_rhyme's elision blindness on hemeling, limerick and rhyme_scheme.

Alternatives considered

Splitting every apostrophe-bearing token at the tokeniser. Rejected: it would misread the 94 genuine apostrophe words the tokeniser was deliberately built to keep whole, and every other row that tokenises French text — not only these two — would inherit the change untested.

Reconstructing correct elision on the output (choosing l'/le/la from the displaced noun's gender and initial sound). Rejected for this decision: the noun list carries no gender, so it would need a second data source and a design of its own, for a cost this fix does not need to pay to close the sweep's finding — the finding was a verdict inversion, not a request for grammatical polish.