Record coarse vision spot-check

vision/quality-audit.json

@@ -1,6 +1,6 @@
 {
   "scope": "Automated structural and image-decoding checks over the materialized Jev vision config.",
-  "manual_review": "The emitted sample rows must be opened with their images to judge visual answerability.",
+  "manual_review": "Coarse visual spot-check completed after publication: one final recast train row per source (14/14 sources) was opened. Images, question/options, and selected gold were coherent with no obvious recast mismatch. This is a spot-check, not exhaustive semantic annotation.",
   "splits": {
     "train": {
       "rows": 5020,
@@ -969,5 +969,7 @@
       "image_group_id": "b86b7a87e58771266f4a5295",
       "manual_review": "pending"
     }
-  ]
+  ],
+  "manual_review_sources": 14,
+  "manual_review_rows": 14
 }