fix: relax ocr noise classification

backend/app/services/review_service.py

@@ -1023,14 +1023,18 @@ class ReviewService:
             return False
         if PDF_PAGE_NOISE_PATTERN.fullmatch(normalized):
             return True
+        lowered = normalized.lower()
+        decoration_cues = ("page ", "pdf ", " pdf", "sample", "footer", "header", "for extraction", "copyright", "appendix")
+        if any(cue in lowered for cue in decoration_cues):
+            return True
+        if re.fullmatch(r"[-_/\\|·•\s\d]+", normalized):
+            return True
         if len(normalized) > 80:
             return False
         word_count = len(re.findall(r"[A-Za-z0-9\u4e00-\u9fff]+", normalized))
-        if word_count > 8:
-            return False
         if any(marker in normalized for marker in PDF_TEXT_MARKERS):
             return False
-        return True
+        return word_count <= 2 and len(normalized) <= 18
 
     def _looks_like_platform_review_report(self, value: str) -> bool:
         marker_hits = sum(1 for marker in REPORT_PDF_MARKERS if marker in value)