Widen score epsilon to 0.01 so :.3f formatting stays in (0, 1)

The validator parses score=X.XXX from the [END] line of inference.py
stdout. With eps=1e-9 and :.3f formatting, scores rounded to "0.000"
or "1.000" which the validator reads as exactly 0.0 or 1.0.

Using eps=0.01 guarantees the formatted score is always strictly
between 0.010 and 0.990.

backend/env/tasks.py

@@ -330,7 +330,7 @@ def get_all_tasks() -> list[Task]:
     return list(TASKS.values())
 
 
-_EPS = 1e-9
+_EPS = 0.01  # wide enough that f"{x:.3f}" never rounds to 0.000 or 1.000
 
 
 def grade_response(

inference.py

@@ -200,11 +200,12 @@ async def run_episode(
             if done:
                 break
 
-        # Score: average of per-step rewards (each already in (0,1) from env)
-        # Clamp to [0, 1] as a safety net
-        total = sum(rewards)
-        max_possible = MAX_STEPS * 1.0
-        score = min(max(total / max_possible if max_possible > 0 else 0.0, 0.0), 1.0)
+        # Score: average of per-step rewards. Clamp strictly inside (0, 1)
+        # with margin >= 0.005 so f"{score:.3f}" never formats to "0.000" or "1.000".
+        _EPS = 0.01
+        denom = max(len(rewards), 1)
+        avg = sum(rewards) / denom if rewards else _EPS
+        score = max(_EPS, min(1.0 - _EPS, avg))
 
     finally:
         log_end(