Guard best_cycles with correctness

app.py

@@ -483,8 +483,9 @@ def perf_takehome_reward_fn(completions, prompts=None, **kwargs):
         if code.strip() in EXAMPLE_CODE_SET:
             reward = max(0.0, reward - COPY_PENALTY)
         cycles = result.get("cycles")
+        correctness = result.get("correctness", 0.0)
         with state_lock:
-            if isinstance(cycles, int) and cycles < training_state["best_cycles"]:
+            if isinstance(cycles, int) and isinstance(correctness, (int, float)) and correctness > 0 and cycles < training_state["best_cycles"]:
                 training_state["best_cycles"] = cycles
                 training_state["best_code"] = code
         rewards.append(float(reward))