abkhaz: ⟨ҩ⟩ is pharyngealized, and an identical glide twin is not its own vowel context by christopherthompson81 · Pull Request #768 · christopherthompson81/vernacula-phonemizer

christopherthompson81 · 2026-08-08T18:46:41Z

What

The two genuine engine defects among the referee's residual divergence classes, sorted from the referee's own inconsistencies in docs/abkhaz_phonology_investigation.md:

⟨ҩ⟩ → [ɥˤ], not plain [ɥ]. The word corpus writes the pharyngealized form ~12:1 — kaikki 50× ɥˤ against 4 bare, and wikipron's one non-definition ҩ-word (ахҩа) agrees. The bare-ɥ rows are the letter definition and the numeral series, which is the referee's demonstrably inconsistent corner (it also devoices б there, against its own б=b definition). Chirikba /ʕʷ/ and Hewitt [ɥˤ] are the same segment. The manifest's stated policy — follow the corpus where the sources disagree — decides it.
Identical glide twins: ⟨у⟩/⟨и⟩ count each other as vowel context, so ⟨уу⟩ made both glides — асууари read as[ww]ari. A twin on the right no longer counts, so the first is syllabic: асууари → asuwari, диит → dijtʼ (was djjtʼ), Russian -ии → ij. Mixed runs (аиуит → ajwjtʼ) unchanged.

Plus two kaikki notation folds (ʏ for ⟨ы⟩ after ɥˤ; e̞).

Verification

Referee: kaikki 641 → 676 (69.1%), raw 389 → 409. Primary 170 → 166 — all four losses are the Ҩ/ҩ definition rows and the bare-ɥ numerals, i.e. the corner the change deliberately overrules; net +31 across both referees.
Corpus: 347/404 rows changed (⟨ҩ⟩ is everywhere). Mechanically verified — strip ˤ from the after side and diff — that every changed row is one of the two intended classes; the 47 residuals are all ⟨ии⟩→[ij]. DROP steady at 33, no leaks.
Documented-and-left (with greps): the referee's phonemic-analysis rows ([i]=əj, [e]=aj), the devoiced numeral series, dropped contrastive ʰ/ʼ in single rows, and the Гь→c-family definitions already rejected in core/ipa: ⟨ɑ ɒ ʏ⟩ join the nucleus class — stress was vanishing, bhojpuri +125 words #752.
npm run ci green (3157 tests). fr/ht keep their plain ɥ; ɥ is in no shared vowel class.

🤖 Generated with Claude Code

https://claude.ai/code/session_012Lc3WnUgogC7okV7n53vjr

… own vowel context The word corpus writes [ɥˤ] for ⟨ҩ⟩ ~12:1 (kaikki 50× vs 4; wikipron agrees outside the definition rows); the bare-ɥ rows are the letter definition and the numeral series — the same inconsistent corner that devoices б. Chirikba /ʕʷ/, Hewitt [ɥˤ]. And ⟨уу⟩/⟨ии⟩ made BOTH letters glides because у/и count each other as vowel context — асууари read as[ww]ari; the first twin is now syllabic (asuwari, dijtʼ), mixed runs unchanged. Two notation folds (kaikki ʏ for ы after ɥˤ, e̞). Referee: kaikki 641 → 676 (69.1%); primary 170 → 166, all four losses in the definition/numeral corner. Corpus: 347/404 rows changed, mechanically verified to be only the two intended classes; DROP steady, no leaks. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012Lc3WnUgogC7okV7n53vjr

…fixes The twin patch was keyed on letter identity, which fixed only uu-after- consonant: ауу still read a[ww] and адиуан read adjwan against the referee own adiwan. The general rule — left context is the realized previous phone, right context excludes an undecided у/и — repairs every nucleus-free glide run: 187/404 corpus rows change, all of them wj→uj / jw→iw (уи "that" had no nucleus; Хьиуитт was χʲjwjtʼtʼ). kaikki 676 → 677. Also: ҩ+apostrophe no longer doubles the pharyngealizer (ɥˤˤ); the dead e̞ fold is removed (BACKBONE strips U+031E first) and ʏ→ə is context-limited to post-ɥˤ; the vowelLetters comment, the referee-floor comment, the catalogue row and the test comments now state the current rule and numbers, with the no-referee pins marked as invariant pins. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012Lc3WnUgogC7okV7n53vjr

…t (and refresh ab) `tools/language-catalogue/derive-normalization.py` decides the `normalization` column from exactly two things — does the engine directory have a `normalize.ts`, and does the engine call it. pcm's treatment (#781) put its pre-tokenizer rewrites INLINE in naija.ts, so the column stayed empty and pcm read as UNTREATED in the planning query the catalogue exists to serve. Typing `done` into the TSV would have been overwritten on the next run. ⚠ Of the 62 engine DIRECTORIES that use the symbol tier (73 registry codes, several sharing a directory), naija was the ONLY one with no normalize.ts — checked, not assumed. So this is pcm being unconventional, not the tool being narrow, and the fix is to be conventional rather than to teach the tool a second pattern it would then have to keep matching. Moves the abbreviation expansion (a text rewrite, the same slot Hausa's `normalizeHausa` occupies) into src/languages/naija/normalize.ts. The SYMBOL tier stays in naija.ts with the engine data, as it does for ha/af/ur. No behaviour change: 3192 tests unchanged, and the call site is still SYMBOLS(normalizeNaija(input)) — the order was deliberate. ⚠ THE .db REBUILD IS NOT A TWO-CELL CHANGE. main's committed languages.db did not match main's own catalogue.tsv: a fresh build differs on 60 rows, staleness accumulated because the TSV has been edited (e.g. #768/#769 rewrote ab's notes) without rebuilding the binary. This commit's .db is byte-identical to a fresh build, so it also flushes those 59 unrelated rows — which is why the binary diff is large for a change described as two cells. Regenerating the column also picked up a second stale cell unrelated to this work: `ab` (Abkhaz) has had a normalize.ts since its own treatment landed and was still reported as untreated. Both are now `done` — 82 → 83, (none) 113 → 112. The `pr` column is left empty, matching ha and ab: it carries a value in only 7 of 212 rows and three of those hold stray note text, so it is not a maintained convention to opt into here. ⚠ The derivation and the .db build are both hand-run, which is how ab and the 60 rows drifted. A `--check` gate in `npm run ci` (the shape check:package already uses) would catch both classes; flagged rather than done here. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01J9yuDGsezBBo5myS3AeN7D

…t (and refresh ab) (#782) `tools/language-catalogue/derive-normalization.py` decides the `normalization` column from exactly two things — does the engine directory have a `normalize.ts`, and does the engine call it. pcm's treatment (#781) put its pre-tokenizer rewrites INLINE in naija.ts, so the column stayed empty and pcm read as UNTREATED in the planning query the catalogue exists to serve. Typing `done` into the TSV would have been overwritten on the next run. ⚠ Of the 62 engine DIRECTORIES that use the symbol tier (73 registry codes, several sharing a directory), naija was the ONLY one with no normalize.ts — checked, not assumed. So this is pcm being unconventional, not the tool being narrow, and the fix is to be conventional rather than to teach the tool a second pattern it would then have to keep matching. Moves the abbreviation expansion (a text rewrite, the same slot Hausa's `normalizeHausa` occupies) into src/languages/naija/normalize.ts. The SYMBOL tier stays in naija.ts with the engine data, as it does for ha/af/ur. No behaviour change: 3192 tests unchanged, and the call site is still SYMBOLS(normalizeNaija(input)) — the order was deliberate. ⚠ THE .db REBUILD IS NOT A TWO-CELL CHANGE. main's committed languages.db did not match main's own catalogue.tsv: a fresh build differs on 60 rows, staleness accumulated because the TSV has been edited (e.g. #768/#769 rewrote ab's notes) without rebuilding the binary. This commit's .db is byte-identical to a fresh build, so it also flushes those 59 unrelated rows — which is why the binary diff is large for a change described as two cells. Regenerating the column also picked up a second stale cell unrelated to this work: `ab` (Abkhaz) has had a normalize.ts since its own treatment landed and was still reported as untreated. Both are now `done` — 82 → 83, (none) 113 → 112. The `pr` column is left empty, matching ha and ab: it carries a value in only 7 of 212 rows and three of those hold stray note text, so it is not a maintained convention to opt into here. ⚠ The derivation and the .db build are both hand-run, which is how ab and the 60 rows drifted. A `--check` gate in `npm run ci` (the shape check:package already uses) would catch both classes; flagged rather than done here. Claude-Session: https://claude.ai/code/session_01J9yuDGsezBBo5myS3AeN7D Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>

christopherthompson81 mentioned this pull request Aug 8, 2026

abkhaz: км/м are words, the dot decimal reads, and the sourcing gate follows a manifest #769

Merged

christopherthompson81 merged commit c38ecbf into main Aug 8, 2026

christopherthompson81 deleted the abkhaz-phonology-residuals branch August 8, 2026 19:24

Navigation Menu

abkhaz: ⟨ҩ⟩ is pharyngealized, and an identical glide twin is not its own vowel context - #768

abkhaz: ⟨ҩ⟩ is pharyngealized, and an identical glide twin is not its own vowel context#768
christopherthompson81 merged 2 commits into
mainfrom
abkhaz-phonology-residuals

christopherthompson81 commented Aug 8, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

christopherthompson81 commented Aug 8, 2026

What

Verification

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant