fix: keep the curly apostrophe when folding a lookalike word

.changeset/keep-curly-apostrophe.md

@@ -0,0 +1,11 @@
+---
+'@moderation-api/unicode-spoofing': patch
+---
+
+Keep the typographic apostrophe in `normalized` when a lookalike word is folded.
+
+`Ivа’s` (with a Cyrillic `а`) normalized to `Iva's`: the fold ran every
+character of the word through the confusables map, including U+2019, whose
+prototype is the ASCII `'`. Only the letters are lookalikes, so the apostrophe
+is now kept as typed and the word normalizes to `Iva’s`. Detection is
+unchanged.

src/analyze.ts

@@ -425,6 +425,9 @@ function analyzeToken(
           return '';
         if (zalgo && MARK_RE.test(ch)) return '';
         if (nfkcFold) return ch;
+        // The apostrophe is punctuation the token admits, not a lookalike:
+        // folding it would turn "Ivа’s" into "Iva's" rather than "Iva’s".
+        if (ch === '’') return ch;
         return foldLookalikes ? foldChar(ch) : ch;
       })
       .join('');

test/analyze.test.ts

@@ -133,6 +133,14 @@ describe('analyze — legitimate content', () => {
     expect(r.signals.mixed_script).toBe(true);
   });
 
+  it('keeps the curly apostrophe when it folds a lookalike word', () => {
+    // Only the letters are lookalikes; the apostrophe is the author's own
+    // punctuation and survives normalization as typed.
+    expect(analyze('Ivа’s').normalized).toBe('Iva’s');
+    expect(analyze("Ivа's").normalized).toBe("Iva's");
+    expect(analyze('the boss’ НОТ’ deal').normalized).toBe('the boss’ HOT’ deal');
+  });
+
   // A word already written in Latin cannot be impersonating Latin, so the only
   // thing the skeleton test measures there is whether the fold happens to reach
   // ASCII — which it does for ordinary European letters, and arbitrarily: "æ"

test/cases.data.ts

@@ -64,6 +64,13 @@ export interface SpoofCase {
 }
 
 export const CASES: SpoofCase[] = [
+  {
+    category: 'confusable',
+    name: 'keeps the curly apostrophe of a folded word (Ivа’s → Iva’s)',
+    input: 'Ivа’s',
+    status: 'supported',
+    expect: { spoofed: true, signals: { mixed_script: true }, normalized: 'Iva’s' },
+  },
   {
     category: 'confusable',
     name: 'flags a Cyrillic whole-word lookalike and normalizes it (НОТ → HOT)',