fix: import json, use make_env_base in phase switch, and run eval sequentially to avoid second concurrent sim car

agent/experiments/exp14_finetune_v5.py

@@ -11,7 +11,7 @@ Exp 14b: Fine-tune mountain champion (v5 reward) — throttle schedule.
 This script is intentionally conservative: small LR, frequent evals, and
 saves so we don't lose progress.
 """
-import sys, os, time
+import sys, os, time, json
 sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..'))
 from datetime import datetime
 from stable_baselines3 import PPO
@@ -197,7 +197,7 @@ try:
                 env0.close()
             except Exception:
                 pass
-            env0 = VecTransposeImage(DummyVecEnv([make_env(throttle_min)]))
+            env0 = VecTransposeImage(DummyVecEnv([make_env_base(0.2, throttle_floor=throttle_min)]))
             model.set_env(env0)
 
         remaining = phase_steps
@@ -213,8 +213,17 @@ try:
             model.save(os.path.join(SAVE_DIR, 'model'))
             log(f'[{steps_done}/{TOTAL_STEPS}] Checkpoint saved: {ckpt}.zip')
 
-            # Eval
+            # Eval sequentially: close training env first so we do not create a
+            # second concurrent sim connection / second car during checkpoint eval.
+            try:
+                env0.close()
+            except Exception:
+                pass
             res = evaluate_model(model, throttle_min, sets=EVAL_EPISODES)
+            # Restore training env after eval
+            env0 = VecTransposeImage(DummyVecEnv([make_env_base(0.2, throttle_floor=throttle_min)]))
+            model.set_env(env0)
+
             # compute mean lap time for episodes that completed at least 1 lap
             lap_times = [lt for r in res for lt in r['lap_times']]
             mean_lap = sum(lap_times)/len(lap_times) if lap_times else None