fix(OpenEnv): fix rounding bug in inference.py log and ensure safe [0.1, 0.9] interior scores

grader.py

@@ -32,8 +32,8 @@ from models import EpisodeResult, ReviewArtifact, ReflectionArtifact
 # ─────────────────────────────────────────────
 # SCORING CONFIG
 # ─────────────────────────────────────────────
-_SCORE_MIN = 0.001   # never exactly 0.0
-_SCORE_MAX = 0.999   # never exactly 1.0
+_SCORE_MIN = 0.10   # deep interior to avoid all boundary issues
+_SCORE_MAX = 0.90   # deep interior to avoid all boundary issues
 
 
 def _clamp(score: float) -> float:
@@ -166,7 +166,6 @@ def score_review_quality(
     code_words   = re.findall(r'\b[a-z_]{3,}\(\)', combined)
     specificity  = min(0.1, len(set(code_words)) * 0.025)
 
-    # Strictly (0, 1) - Safer interior
     return _clamp(kw_score * 0.7 + length_bonus + specificity)
 
 
@@ -318,13 +317,16 @@ def grade_task(repo_path: str, **kwargs) -> float:
         except Exception:
             pass
 
-    result = grade_episode(
-        repo_path=repo_path,
-        task_id=task_id,
-        total_steps=total_steps,
-        max_steps=max_steps,
-        reviews=reviews,
-        reflections=reflections,
-        required_keywords=required_keywords,
-    )
-    return float(result.final_score)
+    try:
+        result = grade_episode(
+            repo_path=repo_path,
+            task_id=task_id,
+            total_steps=total_steps,
+            max_steps=max_steps,
+            reviews=reviews,
+            reflections=reflections,
+            required_keywords=required_keywords,
+        )
+        return float(result.final_score)
+    except Exception:
+        return _SCORE_MIN

inference.py

@@ -237,11 +237,11 @@ def run_episode(env: TeamForgeEnv, agent: Agent, task_id: str) -> Dict:
     score   = result.final_score
     success = result.passed
 
-    rewards_str = ",".join(f"{r:.3f}" for r in rewards) if rewards else "0.001"
+    rewards_str = ",".join(f"{r:.4f}" for r in rewards) if rewards else "0.1000"
 
     # ── [END] ─────────────────────────────────────────────────────────────────
-    # We use 2 decimal places to match common validator expectations, 
-    # but the internal value is strictly interior [0.1, 0.9].
+    # We use 4 decimal places to ensure that interior scores (e.g. 0.999)
+    # are never rounded to illegal boundary values (1.00) in the logs.
     print(
         f"[END] success={'true' if success else 'false'} steps={step_count} "
         f"score={score:.4f} rewards={rewards_str}",

openenv.yaml

@@ -116,7 +116,7 @@ observation_space:
 
 # ── Reward ─────────────────────────────────────────────────────────────────────
 reward:
-  range: [0.001, 0.999]
+  range: [0.0, 1.0]
   type: dense
   description: >
     Dense shaped reward. Positive for: correct plan steps, edits, passing tests,
@@ -129,21 +129,21 @@ tasks:
     max_steps: 20
     description: "Fix an off-by-one bug in utils/list_ops.py. All 7 tests must pass."
     grader: grader.grade_task
-    score_range: [0.001, 0.999]
+    score_range: [0.0, 1.0]
 
   - id: medium_refactor_stats
     difficulty: medium
     max_steps: 30
     description: "Refactor monolithic stats.py into a stats/ package. 15 tests must pass with full backward compatibility."
     grader: grader.grade_task
-    score_range: [0.001, 0.999]
+    score_range: [0.0, 1.0]
 
   - id: hard_lru_cache_performance
     difficulty: hard
     max_steps: 40
     description: "Implement O(1) LRU cache from a stub. 15 correctness tests + 1 performance test (10k ops < 200ms)."
     grader: grader.grade_task
-    score_range: [0.001, 0.999]
+    score_range: [0.0, 1.0]
 
 # ── Infrastructure ─────────────────────────────────────────────────────────────
 runtime: