fix(keyword-matcher): fold Latin diacritics so "PREÇO" matches a "preco" keyword

[reisvini]

The bug

stripSpecialCharacters keeps \p{L}, which is correct, but it means an accented comment and an unaccented keyword stay different strings and never match. Measured against main (c51f04d), these all miss in both whole-word and partial mode:

comment	keyword	result
PREÇO?	preco	miss
ÍNDICE	indice	miss
informação	informacao	miss
¿CUÁL es el precio?	cual	miss

The reverse direction misses too: a campaign keyed on preço never matches a commenter who typed preco.

Accents get typed inconsistently, and mobile keyboards differ on how easy they are to reach, so for Portuguese, Spanish, French and German campaigns this drops real matches. It drops them quietly, because a comment that does not match leaves no log row to notice.

The fix

foldDiacritics normalizes to NFD, drops combining marks, recomposes. Applied to both sides inside matchKeywords, so it is symmetric.

The fold is scoped to Latin-script bases, and that scoping is the substance of this PR rather than caution for its own sake. The usual one-liner:

text.normalize("NFD").replace(/\p{M}/gu, "")

is wrong for a tool that has already gone to the trouble of supporting non-Latin scripts, because a combining mark is load bearing outside Latin. Measured, that one-liner does this:

input	naive fold	what broke
किताब	कतब	Devanagari vowel signs deleted
สวัสดี	สวสด	Thai vowel marks deleted
مَرْحَبًا	مرحبا	Arabic harakat deleted
йод	иод	different Cyrillic word
Україна	Украіна	Ukrainian ї folded to і
ガード	カート	dakuten flipped: "guard" becomes "cart"

Walking the decomposed string and dropping a mark only when the base character before it is \p{Script=Latin} leaves every one of those byte for byte identical. The tests assert that directly, including the Japanese case.

Input is decomposed before the walk so precomposed (U+00E9) and decomposed (U+0065 U+0301) sources behave the same. The Instagram API returns both.

What does not change

• Whole-word boundaries still hold after folding: apreço does not match a preco keyword in whole-word mode.
• Non-Latin keywords match exactly as before, and иод still does not match йод.
• matchedKeyword is still returned exactly as the campaign spelled it, not the folded form, so the DM log and the UI are unaffected.
• stripSpecialCharacters keeps its existing contract and its existing tests. The fold is a separate exported function.

Out of scope on purpose

Ligature folding. Fußball does not match fussball, because ß is a letter rather than a diacritic and folding it needs a locale-aware mapping rather than NFD. Happy to do it in a follow-up if you want it.

Checks

All four from CONTRIBUTING.md pass locally:

• npm run typecheck clean
• npm run lint clean
• npm test 14 files, 158 tests passing (13 added here)
• npm run build succeeds

__tests__/follower-history.test.ts needs npm run db:generate before it will load, on main as well as on this branch. Worth a line in CONTRIBUTING.md if you agree.

Why I was in here

I was reading OpenReply closely while planning comment-to-DM automation for a product that ships pt-BR and es-LA as first-class locales, which is how this surfaced. Thanks for open-sourcing it, the Unicode note already in the matcher header is what made the gap easy to spot.

[vercel]

@reisvini is attempting to deploy a commit to the diwenne's projects Team on Vercel.

A member of the Team first needs to authorize it.