42. The blend needs no grapheme-to-phoneme model, and German needed a fold

Context

ADR 0041 shipped paronomasia with a bound: it checks a displacement only when the word that lands is itself in the pronouncing dictionary. The blends that dominate the punning shopfront — Haarmonie, Hairitage, Föhnix, Chaarisma, Haarchitektur — put a coined word there, and no pack here can pronounce a word it has never seen. That ADR named phonemes.g2p as the capability the family was blocked on, in the manner ADR 0015 prescribes: named on the row, with no pack behind it.

The question then asked was whether to build it. Three routes were costed: espeak-ng as an out-of-tree GPL-3.0 plugin; a model learned from the 837,689 German Wiktionary entries, 134,000 CMUdict entries and 125,343 Lexique entries already shipped; or nothing.

Measurement first, and it changed the question. A blend is not an arbitrary string. The host word and the word spliced into it are both ordinary dictionary entries, so a blend can be verified as a derived form — the host recoverable behind it, the spliced word really present, and that word sounding like a stretch of the host — without ever pronouncing the coinage. Measured against the named targets, using nothing that is not already installed:

coinage	host	spliced	distance
Föhnix	Phönix	Föhn	0.000
Hairitage	heritage	hair	0.000
Breadwinner	breadwinner	bread	0.000
Hairways	airways	hair	0.333
Haarmonie	Harmonie	Haar	0.667
Chaarisma	Charisma	Haar	0.667

English worked immediately. German did not, and the reason was not g2p at all: German vocalises a syllable-final /r/, so Haar is transcribed h aː ɐ̯ while Harmonie opens h a ʁ. The same underlying segment, spelled two ways, and an unweighted edit distance charges a full substitution for it — the cost ADR 0041 D3 admitted when it chose unweighted Levenshtein over a feature-weighted metric.

Decision

D1. No phonemes.g2p, and the name is retired rather than left standing. The family it was named for is reachable without it. What remains out of reach is not g2p-shaped: a host the lexicon does not carry (usually a proper noun — Barbarella is absent from CMUdict) and the cross-lingual blend (Atmosph'air), which needs phonemes.bilingual and would need it whatever g2p existed.

Recorded as a decision because the alternative was attractive and expensive. espeak-ng costs a system binary, a GPL-3.0 quarantine and a distribution of its own, and its output would still need mapping into each pack's notation — which for English is ARPABET, so the mismatch is real work, not a wrapper. A learned model costs no licence, but it is a statistical component in a project whose whole claim is that its verdicts are decidable: at 90% accuracy one verdict in ten is wrong for a reason nobody can see, which is a far worse failure than today's MissingCapability, and it would have to be evaluated against the very dictionaries it was trained on.

D2. portmanteau is its own row. paronomasia displaces a whole word; a blend splices inside one. Different operations, different violations, and ADR 0007 is one module per procedure. Folding both into one row would have made its catalogue definition describe two things, which is the drift ADR 0015 exists to prevent.

D3. A per-language equivalence table, and it folds one thing. core.phonetics.EQUIVALENT records symbols that are one segment in a language's own notation. German gets {ʁ, ɐ̯, ɐ, r}. Vowel length is deliberately not folded: aː and a are contrastive — Staat and Stadt — so collapsing them would buy Haarmonie a distance of 0.000 by discarding a distinction the language makes. It sits at 0.333 instead. English and French get no entries: ARPABET has no vocalised-r symbol and Lexique's SAMPA maps to a single ʁ, so a table for them would be a guess dressed as data.

D4. The window is searched, not aligned, and that is the row's weakest joint. Deciding which phonemes of the host the spliced letters cover would need grapheme-to-phoneme alignment, which this project does not have. The check asks instead whether the spliced word sounds like some window of the host of about the right length. That would accept a blend whose word matches a stretch somewhere other than where it was spliced. It is the strongest reading available without an aligner, and it is stated in the module docstring rather than left for a reader to discover.

D5. apply is not shipped, though kind says both. The figure is plainly constructive, so misdescribing it as restrictive to match the implementation would be the wrong direction. Generating a blend means choosing where to splice — Deri and Knight's multitape-FST problem — and is its own piece of work. describe_procedure will still report both; the docstring heads that off, as larding and lipogrammatic_translation already do.

Consequences

Folding was verified additive before it shipped. All nine German pairs the existing fixtures and domain phrases turn on — Kamm/Komm, Welle/Welt, Zopf/Kopf, Glanz/Ganz and the rest — score identically with and without it, and a test pins that. So this is not a re-reading of work already checked.

The prefilter had to learn the same fold, and briefly did not. shares_an_edge compares first and last phonemes to decide what is worth measuring; unfolded, it rejected Haar against Harmonie before the metric ever saw the pair. A prefilter stricter than the measure behind it hides candidates the band would have accepted, silently. It now folds too.

ADR 0041's phonemes.g2p note is now wrong and stays wrong on purpose. The row's notes and module docstring named it as the blocker in good faith and the measurement overturned it; paronomasia's notes now point here instead. The older ADR is a record of what was decided when, not a document to keep tidy.

A coinage that is its own host is not a blend, and the data found that before a reader would have. Crustacean contains crust; Breadwinner contains bread; both satisfied an early draft at a spelling distance of 0.000, because nothing had been spliced. That is the found pun, which paronomasia already excludes on the grounds that it states no relation between two texts. The guard is identical_to_host, and it exists because three trades' worth of authored blends were run through the checker rather than admired.

Two things are still unreachable, and neither is g2p's fault. A proper-noun host, which is most of the film and celebrity blends; and the cross-lingual blend, which is a large share of the German salon window in practice — British Hairways over a Berlin shop is English inside German, and phonemes.bilingual is what that needs.

Coverage moves. 123 implemented rows become 124 and the catalogue 157 becomes 158; eight constructed cases take the corpus from 572 to 580. No case claims external provenance, for ADR 0041 D7's reason, which has not changed.