Refine prompt_template_source: cite verifiable official sources per 2026-05-30 deep audit

metadata.json

@@ -47,9 +47,9 @@
         }
       },
       "prompt_template_source": {
-        "origin": "fallback",
-        "reference": "T2-adapted",
-        "notes": "single-image per-dataset adaptation"
+        "origin": "official",
+        "reference": "https://github.com/EvolvingLMMs-Lab/lmms-eval/blob/main/lmms_eval/tasks/winoground/utils.py (winoground_doc_to_text — two-image visio-linguistic compositional reasoning)",
+        "notes": "lmms-eval Winoground canonical evaluation prompt."
       }
     }
   }