mai: port Maithili to C#, and pin the divergences the golden cannot see

[christopherthompson81]

Maithili runs on the shared makeNativeHindi abugida engine, so the port is 39 lines of
data-wiring plus one thing no sibling has: a normalize override that folds ⟨॑⟩ U+0951 onto
the avagraha ⟨ऽ⟩. 200/200 first run; the full gate goes 82 → 83 languages, 16,400 → 16,600
rows, 0 differ. C# tests 732 → 752, vitest 5,408 → 5,412.

⚠ THE GOLDEN IS INERT ON THE MODULE'S SIGNATURE. Measured over its 200 mined rows: ⟨े⟩ 1,251,
⟨ो⟩ 385, ⟨ै⟩ 170, ⟨ौ⟩ 47 — the vowel divergences are exercised hard — but ⟨॑⟩ is 9 occurrences
in 2 rows and EVERY ONE is a monosyllable, where retainInMonosyllable already keeps the vowel
and the reading is identical fold or no fold. ⟨ऎ ऒ ॆ ꣿ ऋ ₹ ॐ ळ⟩ are 0/200 and no golden row
carries Latin. Off-golden: the whole mined pool + the attest examples + 130 adversarial lines
= 535 lines sync and async, and 6,408 word types through phonemizeWord, all 0 differ.

Fixed TypeScript-first, no golden row moved:

• phonemizeWord did not see the fold. word() does not run the normalizer — the override
  reaches text() only — so the path the referee eval scores read अब॑ as ˈəb while the
  shipped path read ˈəbə. The eval and the product were different engines on exactly the
  construct this module exists for, and no golden could ever have shown it.
• The fold's own comment was wrong about why it sits where it does. It claimed the word
  tokenizer "does not carry U+0951 in its class"; DEVANAGARI_WORD is ऀ-ॣॲ-ॿ, which spans
  U+0951, and phonemize("म॑थिली", "hi") — no fold — reads one word.
• maithili.jsonc repeated bhojpuri.jsonc's ₹ claim verbatim, trailing half-sentence and all:
  mai declares no symbolTier, so the shared Hindi tier claims the sign and ₹500 reads
  pˈaː̃t͡ʃ sˈəʊ ɾˈʊpje, not the bare number.
• ⟨ꣿ⟩ U+A8FF is a declared vowelSigns entry the engine cannot reach. Devanagari Extended is
  outside the shared word class, so the mark ENDS the token: कꣿ → kˈə and मꣿथिली → two
  words. Annotated rather than deleted — the alternative fix widens a class that reaches all 17
  makeNativeHindi languages, and U+A8FF is ×0 in the corpus anyway.

Of the three findings the bho port filed against the shared Devanagari data, none recurs:
⟨ऋ⟩/⟨ृ⟩ are the TAP ɾɪ here, matching the manifest's only rhotic; ⟨अ⟩ and inherentVowel are both
ə; and बजकर/मिनट ARE attested for Maithili by the per-sentence morphology test. प्रतिशत is the
one word that is not, and the count is now recorded where it is used: % occurs 13 times across
7 of the 200 golden rows.

Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01WuH4pRrB1PAEiG8CCLVR5S

Also found, NOT fixed here — the ₹ claim is FAMILY-WIDE

grep -rn 'पाँच सौ), matching' data/languages/*/*.jsonc returns five manifests carrying the
identical wrong comment (the same text #1029 removed from bhojpuri.jsonc), including the trailing
half-sentence. Only maithili.jsonc is fixed in this PR, because three of the others are being
ported concurrently in other worktrees:

manifest	line	phonemize("₹500", …)	verdict
maithili.jsonc	303	pˈaː̃t͡ʃ sˈəʊ ɾˈʊpje	false — fixed here
awadhi.jsonc	313	pˈaː̃t͡ʃ sˈʌu ɾˈʊpjeː	false
chhattisgarhi.jsonc	296	pˈaː̃t͡ʃ sˈɔː ɾˈʊpjeː	false
magahi.jsonc	311	pˈɑ̃t͡ʃ sˈɔ ɾˈupd͡ʒe	false (nobody is porting mag)
nepali.jsonc	194	pˈãt͡s sˈʌj	TRUE — ne really does read the bare number, so its copy of the comment is correct and must be left alone

That last row is the reason this is worth a table rather than a sweep: the same sentence is right in
one manifest and wrong in four.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WuH4pRrB1PAEiG8CCLVR5S