feat: optimize padding and formula fallback

app/services/ocr_service.py

@@ -890,8 +890,14 @@ class GLMOCREndToEndService(OCRServiceBase):
 
         # 3. OCR: per-region (parallel) or full-image fallback
         if not layout_info.regions:
-            raw_content = self._call_vllm(padded, _DEFAULT_PROMPT)
-            markdown_content = self._formatter._clean_content(raw_content)
+            # No layout detected → assume it's a formula, use formula recognition
+            logger.info("No layout regions detected, treating image as formula")
+            raw_content = self._call_vllm(padded, _TASK_PROMPTS["formula"])
+            # Format as display formula markdown
+            formatted_content = raw_content.strip()
+            if not (formatted_content.startswith("$$") and formatted_content.endswith("$$")):
+                formatted_content = f"$$\n{formatted_content}\n$$"
+            markdown_content = formatted_content
         else:
             # Build task list for non-figure regions
             tasks = []