fix(model): parse JSON after thinking output

Handle NuExtract3's documented thinking response format by decoding the final answer after the closing think tag before JSON parsing. Non-thinking extraction remains unchanged and malformed thinking output still fails at its source.

model.py

@@ -53,12 +53,14 @@ def extract(image, text, template, enable_thinking):
             do_sample=False,
         )
 
+    output = processor.batch_decode(
+        generated_ids[:, inputs.input_ids.shape[1] :],
+        skip_special_tokens=True,
+        clean_up_tokenization_spaces=False,
+    )[0].strip()
+
     return json.loads(
-        processor.batch_decode(
-            generated_ids[:, inputs.input_ids.shape[1] :],
-            skip_special_tokens=True,
-            clean_up_tokenization_spaces=False,
-        )[0].strip()
+        output.split("</think>", 1)[1].strip() if enable_thinking else output
     )