fix: fold Latin diacritics so accented comments match unaccented keywords

[dchaudhari7177]

Closes #20

Reproduced first on master:

MISS   "preco"  in "qual o PREÇO?"
MATCH  "preço"  in "qual o preço"

The change

foldLatinDiacritics applied at the top of normalizeCommentText. As you noted, that function already normalises both the keyword and the comment, so one change covers both sides.

text.normalize('NFD').replace(/(?<=[A-Za-z])\p{Diacritic}/gu, '').normalize('NFC')

The scope you asked for falls out of NFD directly: decomposition places the base character immediately before its combining mark, so "only drop a mark whose base is Basic Latin" is a single lookbehind. No script table to keep in sync.

I added the trailing NFC so the function's output stays canonical rather than half-decomposed. Matching would work without it — both sides run through the same normaliser — but it means every untouched script comes out byte-identical to its input, which is a much easier property to assert than "equivalent under some normalisation form".

Acceptance criteria

criterion	result
preco matches PREÇO, preço, PRECO	✅ all three
cafe matches café; uber matches über	✅
й is not folded to и	✅ normalizeCommentText('й') !== normalizeCommentText('и')
が is not folded to か	✅ same assertion
Devanagari and Thai survive unchanged	✅ byte-identical, along with Arabic, Korean and Cyrillic

Two extra cases worth having: preço also matches an unaccented preco comment, since the fold is symmetric; and İstanbul (U+0130, a precomposed capital with dot above) folds correctly, because NFD decomposes it before the lowercase step.

Tests

A Latin diacritic folding block in scripts/test-unit.ts: the eight matching cases as a table, a dedicated test each for the Cyrillic breve and the Japanese dakuten asserting both non-equality and non-matching, a loop asserting six non-Latin strings survive normalisation byte-for-byte, and one guarding that folding widens what counts as equal without widening what counts as a boundary — cafe still does not match cafeteria.

Reverting only src/match.ts while keeping the tests fails 7 of the 42.

Verification

npm test — 42 pass, 0 fail. npm run typecheck — clean. No new dependency.

Note this branches from master independently of #56 (issue #19); both touch src/match.ts but in different functions, so whichever lands first, the second is a clean rebase.