Enforce exclusive score bounds in grader and inference

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

inference.py

@@ -297,12 +297,16 @@ def run_task(llm: OpenAI, task_id: str) -> dict:
                 break
 
         score = obs_dict.get("final_score") or 0.0
+        if score <= 0.0:
+            score = 0.001
+        elif score >= 1.0:
+            score = 0.999
         success = score >= 0.5
 
     except Exception as e:
         print(f"[DEBUG] Task {task_id} failed with error: {e}", flush=True)
         success = False
-        score = 0.0
+        score = 0.001
     finally:
         log_end(success, step, score, rewards)
 

server/releaseops_environment.py

@@ -881,6 +881,11 @@ class ReleaseOpsEnvironment(Environment):
             + 0.10 * efficiency
         )
         score = max(0.0, min(1.0, raw_score - forbidden_penalty))
+        # Hackathon validator requires strict bounds: 0 < score < 1
+        if score <= 0.0:
+            score = 0.001
+        elif score >= 1.0:
+            score = 0.999
 
         return {
             "score": round(score, 3),