Merge origin/main into training/unsloth (sync latest upstream)

.dockerignore

@@ -5,8 +5,6 @@
 .env.*
 
 # Large local artifacts and model weights (not needed by runtime image build).
-Models/
-outputs/
 training_corpus/
 
 code-review-env/Models/

code-review-env/server/app.py

@@ -64,7 +64,9 @@ class TaskRunResponse(BaseModel):
 
 	episode_id: str
 	total_steps: int
+	raw_total_reward: float
 	total_reward: float
+	score: float
 	modules_total: int
 	modules_completed: int
 	done: bool
@@ -224,6 +226,18 @@ UI_INDEX_PATH = Path(__file__).resolve().parent / "static" / "index.html"
 STATIC_ROOT = Path(__file__).resolve().parent / "static"
 
 
+def _strict_score_from_reward(total_reward: float, total_steps: int) -> float:
+	# Map average per-step reward to (0, 1) and avoid exact endpoints required by validator.
+	eps = 1e-6
+	avg_reward = total_reward / max(total_steps, 1)
+	score = (avg_reward + 1.0) / 2.0
+	if score <= eps:
+		return eps
+	if score >= 1.0 - eps:
+		return 1.0 - eps
+	return score
+
+
 def _default_source_root() -> str:
 	configured = os.getenv("GRAPHREVIEW_SOURCE_ROOT", "sample_project")
 	return str(Path(configured).resolve())
@@ -544,10 +558,13 @@ def run_task(task_id: str, payload: TaskRunRequest) -> TaskRunResponse:
 		if done:
 			break
 
+	normalized_score = _strict_score_from_reward(total_reward=total_reward, total_steps=total_steps)
 	return TaskRunResponse(
 		episode_id=episode_id,
 		total_steps=total_steps,
-		total_reward=total_reward,
+		raw_total_reward=total_reward,
+		total_reward=normalized_score,
+		score=normalized_score,
 		modules_total=modules_total,
 		modules_completed=len(completed_modules),
 		done=done,