feat: eye filter flag + QA fixes (Gemma4 prompt, YOLO boundary, PaliGemma score, GDINO skip)

scripts/qa/judges/paligemma.py

@@ -32,9 +32,20 @@ def score(frames, prompt):
         descriptions.append(desc)
     
     combined = " | ".join(descriptions)
+    
+    # Simple text-to-score: check if description mentions key terms from prompt
+    prompt_lower = prompt.lower()
+    desc_lower = combined.lower()
+    score = 50  # default
+    # Boost if prompt elements found in description
+    for word in prompt_lower.split():
+        if len(word) > 3 and word in desc_lower:
+            score += 10
+    score = min(100, score)
+    
     return {
         "agent": "PaliGemma",
-        "score": None,  # raw text, scored later by Gemma4
+        "score": score,
         "reasoning": combined,
         "details": {"descriptions": descriptions}
     }