core/ipa: ⟨ɑ ɒ ʏ⟩ join the nucleus class — stress was vanishing, bhojpuri +125 words

[christopherthompson81]

The last residual from #748 — the one I twice said was blocked. It wasn't, and both the diagnosis and the evidence turned out to be available all along.

What I kept missing

I deferred this on the grounds that the ur/bho referees carry no stress marks, so stress PLACEMENT can't be adjudicated. True — but placement was never the only thing at stake. I never asked what happens to a word with no visible nucleus at all:

applyWeightStress("kaːm")    → kˈaːm        applyWeightStress("kɑːm")    → kɑːm
applyWeightStress("saːlaːɾ") → saːlˈaːɾ     applyWeightStress("sɑːlɑːɾ") → sɑːlɑːɾ

The weight rule's fallback chain ends at "first syllable" — but with no nucleus visible there are no syllables, so it emits nothing. In the live engine: آپ "you" → ɑːp, کام "work" → kɑːm, بازار → bɑːzɑːɾ. Everyday words, silently unstressed. No referee is needed to call that wrong.

And the referee did adjudicate — through the other language

referee	main	with ⟨ɑ ɒ ʏ⟩
ur.wikipron-urd-broad	4382/7709 (56.8%)	4382/7709 — identical
ur.cle-speech	3364/5667 (59.4%)	3364/5667 — identical
bho.grammar-mined	1008/1623 (62.1%)	1133/1623 (69.8%)

+125 words. Urdu really is stress-folded and can't see the change; Bhojpuri can, because IPA_VOWELS is not only the stress nucleus — the Indic engines build SEGMENTAL regexes from it too, so an invisible ⟨ɑ⟩ was corrupting more than stress. The evidence I said didn't exist was one command away, in the corpus I'd already been told was in-repo.

The golden changes

18 updated, and they split:

• 8 are stress appearing where there was none — ɑːbɑːd̪→ɑːbˈɑːd̪, ɦɑːt̪ʰ→ɦˈɑːt̪ʰ, ɑːʋɑːz→ɑːʋˈɑːz, nɑːm→nˈɑːm, ɦzɑːɾ→ɦzˈɑːɾ, mʋɑːd̪→mʋˈɑːd̪, sɑːʈʰʋɑː̃→sˈɑːʈʰʋɑː̃, mʊɾəbbˈɑːʔ.
• The rest are placement moving to what the documented rule (rightmost superheavy → rightmost non-final heavy → first) gives once it can see every syllable. Not referee-confirmed for Urdu; the Bhojpuri number is the evidence that the corrected parse is better overall.

A new test pins the bare-word case directly, so the actual bug can't come back quietly.

Also

IPA_VOWELS and IPA_VOWEL collapse back into one list. The split #748 introduced was a workaround for a question that had an answer.

npm run ci green — 228 files / 3137 tests. docs/ipa_classes_investigation.md Run 9 records the miss: I scoped the question to the first thing I noticed and treated one corpus's silence as the whole evidence base.

🤖 Generated with Claude Code

https://claude.ai/code/session_012Lc3WnUgogC7okV7n53vjr

---

Review addendum

Independent measurement confirmed the blast radius: gu/mr/pa/hi change 0 words, sd changes 18 (stress-only), ur changes 2550 — all stress-only, zero segmental, so the claim that the Urdu referees can't see it is exact. Every referee score is identical except bho.

The bho +125 breaks down as 135 fixed / 10 broken, and the 10 are one class: words written with the avagraha ⟨ऽ⟩ (U+0973), which appears in neither the Bhojpuri nor Hindi engine and is dropped — the invisible ⟨ɑ⟩ had been accidentally suppressing their final-schwa deletion. Removing the accident exposed a pre-existing gap rather than creating one (avagraha matches 15/31 → 5/31). Scope is 31 bho + 2 mai words; nothing else in any referee carries the character. Recorded in the investigation doc and left as its own change, since schwaDeletion in the manifest is already the hook for it.

Also confirmed: no residual "no stress at all" outputs remain except genuinely vowelless tokens (ء→ʔ, ળ→ɭ), and IPA_VOWELS is now the complete IPA vowel chart with no regex metacharacters, so [${IPA_VOWELS}] stays safe to interpolate.