fix: parse multipart/mixed response from /understand (was expecting JSON, got multipart)

app.py

@@ -151,36 +151,46 @@ def _caption_via_understand(audio_path, timeout=600, cancel_check=None):
         logger.warning("[Caption] %s: /understand -> %s (%.0fs)", fname, status, elapsed)
         return None
 
-    # Result may be in the poll response itself or need separate fetch
-    result = None
-
-    # Check if poll response contains the result directly
-    if poll_data and isinstance(poll_data, dict):
-        if poll_data.get("caption"):
-            result = poll_data
-        elif poll_data.get("result") and isinstance(poll_data["result"], dict):
-            result = poll_data["result"]
-
-    # Fallback: try fetching result separately
-    if not result:
-        try:
-            r = _fetch_result(job_id, timeout=120)
-            if r.status_code == 200 and r.text.strip():
-                data = r.json()
-                if isinstance(data, dict):
-                    if data.get("caption"):
-                        result = data
-                    elif data.get("result") and isinstance(data["result"], dict):
-                        result = data["result"]
-        except Exception as exc:
-            logger.warning("[Caption] %s: result fetch failed: %s", fname, exc)
+    # Fetch result — /understand returns multipart/mixed (JSON + latents)
+    try:
+        r = _fetch_result(job_id, timeout=120)
+        if r.status_code != 200:
+            logger.warning("[Caption] %s: result fetch HTTP %d", fname, r.status_code)
+            return None
+
+        content_type = r.headers.get("Content-Type", "")
+
+        # multipart/mixed: extract JSON part (caption metadata)
+        if "multipart" in content_type:
+            boundary = None
+            for part in content_type.split(";"):
+                part = part.strip()
+                if part.startswith("boundary="):
+                    boundary = part.split("=", 1)[1].strip('"')
+            if boundary:
+                import re
+                parts = r.content.split(f"--{boundary}".encode())
+                for part in parts:
+                    if b"application/json" in part:
+                        json_start = part.find(b"{")
+                        json_end = part.rfind(b"}") + 1
+                        if json_start >= 0 and json_end > json_start:
+                            data = json.loads(part[json_start:json_end])
+                            if isinstance(data, dict) and data.get("caption"):
+                                logger.info("[Caption] %s: got caption (%d chars)",
+                                            fname, len(data["caption"]))
+                                return data
+
+        # Plain JSON fallback
+        if r.text.strip():
+            data = r.json()
+            if isinstance(data, dict) and data.get("caption"):
+                return data
 
-    if result:
-        logger.info("[Caption] %s: got caption (%d chars)", fname, len(result.get("caption", "")))
-        return result
+    except Exception as exc:
+        logger.warning("[Caption] %s: result parse failed: %s", fname, exc)
 
-    logger.warning("[Caption] %s: no caption in response. poll_keys=%s",
-                   fname, list(poll_data.keys()) if isinstance(poll_data, dict) else "N/A")
+    logger.warning("[Caption] %s: no caption extracted from result", fname)
     return None