fix: normalize [END] score to (0,1) range — divide by 4.8 max score, clamp to (0.001, 0.999)

inference.py

@@ -145,17 +145,22 @@ def run_task(task_name: str) -> dict:
         print(f"  REWARD  : {summary['total_reward']:.3f} / 4.8 max")
         print(f"{'='*60}\n")
 
-        # Required structured block
+        # Required structured block — score must be strictly in (0, 1)
+        MAX_SCORE = 4.8  # 4 steps × 1.2 max reward each
+        raw_score = summary['total_reward']
+        normalized = raw_score / MAX_SCORE
+        # Clamp strictly between 0 and 1 (not 0.0, not 1.0)
+        final_score = max(0.001, min(0.999, normalized))
         print(
-            f"[END] task={task_name} score={summary['total_reward']:.3f} steps={steps_taken}",
+            f"[END] task={task_name} score={final_score:.4f} steps={steps_taken}",
             flush=True,
         )
         return summary
 
     except Exception as exc:
         print(f"[ERROR] run_task({task_name}) failed: {exc}", flush=True)
-        # Still emit END block so validator can parse something
-        print(f"[END] task={task_name} score=0.0 steps=0", flush=True)
+        # Emit END block with minimum valid score (strictly > 0)
+        print(f"[END] task={task_name} score=0.001 steps=0", flush=True)
         return {"task_id": task_name, "status": "error", "total_reward": 0.0, "wrong_steps": 0, "fail_reason": str(exc), "steps": []}