Fix reward.value to always be task_score in (0,1) exclusive

The validator checks reward.value from /step responses. Previously
this was the RL reward (compute_reward) which could be exactly 1.0
on first-attempt success or negative on failure.

Now reward.value = task_score (clamped to (eps, 1-eps) in grade_response).
The RL reward is preserved in info.rl_reward for bandit training.
Also corrects reward_range in /env/info to [0.0, 1.0].

backend/api/openenv.py

@@ -162,7 +162,7 @@ async def env_info():
                 "task_difficulty",
             ],
         },
-        "reward_range": [-1.5, 1.5],
+        "reward_range": [0.0, 1.0],
         "max_steps": 5,
         "tasks": ["simple_queries", "join_queries", "complex_queries"],
         "rl_algorithm": "LinUCB (contextual bandit)",

backend/env/sql_env.py

@@ -332,11 +332,12 @@ class SQLAgentEnv:
 
         obs = self._build_observation()
         reward_info = RewardInfo(
-            value=grader_out.reward,
+            value=task_score,  # always in (eps, 1-eps) per OpenEnv spec
             success=success,
             done=done,
             info={
                 "task_score": task_score,
+                "rl_reward": grader_out.reward,
                 "attempt": ep.attempt_number,
                 "breakdown": {
                     "base": grader_out.breakdown.base,

inference.py

@@ -200,12 +200,11 @@ async def run_episode(
             if done:
                 break
 
-        # Score: clamp sum of rewards to (0, 1) exclusive per OpenEnv spec
-        _EPS = 1e-9
+        # 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  # max reward per step is 1.0
-        raw_score = total / max_possible if max_possible > 0 else _EPS
-        score = max(_EPS, min(1.0 - _EPS, raw_score))
+        max_possible = MAX_STEPS * 1.0
+        score = min(max(total / max_possible if max_possible > 0 else 0.0, 0.0), 1.0)
 
     finally:
         log_end(