manifests: the consonant classes are inventories, not copies — 14 engines

[christopherthompson81]

The third residual from #748. The plan said "consolidate the 33 consonant classes into core/ipa.ts". The triage says don't — and that's the main content of this PR.

Why not

No two of them are the same list. voiceless is six different sets, nasal four, velar five — they share a class name and nothing else:

class	engines	members
voiceless	6	hungarian p t k f s ʃ t͡s t͡ʃ c x · icelandic p t k s θ f h c · luxembourgish p t k s ʃ f χ t͡s t͡ʃ · occitan p t k f s · romanian (obstruent, both voicings) · tibetan (Wylie tokens, not IPA)
velar	5	ancientgreek ɡ k kʰ ks · galician k ɡ · latin k ɡ kʷ ɡʷ kʰ · totontepecmixe k ɡ ɣ · mindong (romanization finals)
sonorant	5	chuvash l r · irish ɾˠ ɾʲ l̪ˠ lʲ · latin l ɫ r · scottishgaelic l̪ˠ lʲ rˠ ɾʲ · tibetan (Wylie)
nasal	4	albanian m n ɲ ŋ · chuvash n m ŋ j · totontepecmixe m n · turkmen m n ŋ

That contrast with the vowels is the whole point: 31 engines held near-identical copies of one universal alphabet, so sharing was obviously right. Here each list is the language's own inventory.

And the narrowness is load-bearing, which is what settles it rather than taste:

• totontepecmixe NASAL is {m, n} — but the engine does emit ŋ. Post-nasal voicing fires after ⟨m n⟩ only.
• occitan VOICELESS_PH excludes ʃ — but the engine does emit it. Intervocalic ⟨s⟩→[z] depends on the exclusion.

A universal class changes both. "The IPA vowel letters" is a fact about the notation; "the voiceless obstruents of this language" is a fact about the language.

What this PR does instead

Moves them to their manifests (#746's rule), each carrying the rule it serves and why it is narrow. core/ipa.ts stays a vowel-class module.

14 engines: albanian, ancientgreek, chuvash, dutch, galician, hungarian, irish, luxembourgish, occitan, romanian, santali, scottishgaelic, totontepecmixe, turkmen.

Irish and Scottish Gaelic are the clearest illustration — their liquids carry the broad ⟨ˠ⟩ / slender ⟨ʲ⟩ marks, so "the liquids" is four phones in a way no shared class could express. Galician already had nasals in its manifest; its velars now sits beside it.

Left alone, deliberately

• tibetan — Wylie stack-grammar classes, not IPA at all; Give the 26 manifest-less engines a manifest, and lift their data out of the .ts #741 kept those with the parser.
• bulgarian / maltese / macedonian VOICELESS — derived from each manifest's DEVOICE map, so already single-sourced.
• icelandic — its files are in flight in icelandic: the hiatus glide is the long í's alone — 8086 → 8091/10093 #749.
• latin's VELAR/LIQUID/MUTA — function-local consts; hoisting them is a separate shape of change.

Verification

Pure data relocation: 19 rewritten sets compared against main, 0 differ (the 20th, Scottish Gaelic's, checked by hand — the script binds DEF/MANIFEST, not that file's M). npm run ci green — 228 files / 3135 tests, typecheck, fence.

docs/ipa_classes_investigation.md Run 8 has the triage.

🤖 Generated with Claude Code

https://claude.ai/code/session_012Lc3WnUgogC7okV7n53vjr