feat(postprocess): normalize curly quotes, ellipsis, NBSP to ASCII

lib/humanizer.ts

@@ -5,7 +5,7 @@ import { getSystemPrompt, getRehumanizePrompt, getCorpusAwareSystemPrompt } from
 import { getCorpusCalibratedThresholds, hasStyleModel, loadStyleModelAsync } from './style-model';
 import { generateWithProvider, getProvider, generateAlternatives } from './providers';
 import { detectAI } from './detector';
-import { postprocess, corpusAwarePostprocess } from './postprocess';
+import { postprocess, corpusAwarePostprocess, normalizePunctuation } from './postprocess';
 import { chunkText, countWords, addToHistory } from './storage';
 import { extractRegions, restoreRegions, containsPlaceholders } from './protect-regions';
 
@@ -106,10 +106,16 @@ export async function humanizeText(
   const targetScore = options.targetScore || calibratedThresholds?.targetScore || 80;
   const maxPasses = options.level === 'ninja' ? 2 : options.level === 'aggressive' ? 2 : 1;
 
+  // Normalize typographic punctuation up front so curly quotes / ellipsis from
+  // the original input (often pasted from Word/Pages/macOS keyboards) don't
+  // survive inside protected regions. Code spans/fences are normalized too, but
+  // the substitutions only affect quotes, ellipsis, and NBSP, not code syntax.
+  const normalizedInput = normalizePunctuation(text);
+
   // Extract structural regions (code, links, URLs, mentions, blockquotes, ...)
   // before any processing. They survive the pipeline as opaque placeholders and
   // are restored verbatim at the end.
-  const { masked, regions } = extractRegions(text);
+  const { masked, regions } = extractRegions(normalizedInput);
   const chunks = chunkText(masked, 2500);
 
   let humanizedText = '';

lib/postprocess.ts

@@ -196,6 +196,23 @@ function stripAIDashes(text: string): string {
     .replace(new RegExp(RANGE_PLACEHOLDER, 'g'), '–');
 }
 
+// ==================== 2b''. NORMALIZE TYPOGRAPHIC PUNCTUATION ====================
+
+// Curly quotes, smart apostrophes, and horizontal ellipsis are another common AI
+// tell (and also a copy-paste hazard from Word/Pages/macOS keyboards). Normalize
+// to ASCII so the body of the rewrite uses only `'`, `"`, and `...`. Code spans
+// and fences are preserved upstream by the protect-regions pass; this only ever
+// runs on prose.
+export function normalizePunctuation(text: string): string {
+  return text
+    .replace(/[‘’‚‛]/g, "'")   // curly singles, low-9, high-reversed-9
+    .replace(/[“”„‟]/g, '"')   // curly doubles, low-9, high-reversed-9
+    .replace(/[‹›]/g, "'")               // single guillemets
+    .replace(/[«»]/g, '"')               // double guillemets
+    .replace(/…/g, '...')                     // horizontal ellipsis
+    .replace(/ /g, ' ');                      // non-breaking space
+}
+
 // ==================== 2c. PUNCTUATION & FORMATTING NOISE ====================
 
 const CONTRACTIONS: [string, string][] = [
@@ -700,6 +717,10 @@ export function postprocess(text: string, options?: PostProcessOptions): string
   const style = options?.style;
   let result = text;
 
+  // ALWAYS: Normalize typographic punctuation (curly quotes, ellipsis, NBSP).
+  // Done early so downstream steps see ASCII-only chars.
+  result = normalizePunctuation(result);
+
   // ALWAYS: Strip em-dashes (strongest AI tell). Numeric ranges preserved.
   result = stripAIDashes(result);