tune(PPO_Bot): logStd=-2.0 (std≈0.135), entropy=0, ceiling=-1.0

Stochastic eval at std≈0.37 was 0/10 vs Corners (deterministic: 10/10).
Warm-start policy is correct but brittle — any noise breaks it.
- log_std initialized to -2.0 (std≈0.135) for moderate exploration
- entropy_coeff=0.0 (no push toward exploration during fine-tuning)
- logStd ceiling=-1.0 (cap at std≈0.37)

tools/training_runner/training.env

@@ -4,7 +4,7 @@ PPOB_LOG_FILE=/home/davide/Projects/SirRoboGarage/tools/training_runner/logs/fir
 PPOB_LR=1e-4
 PPOB_UPDATE_INTERVAL=10
 PPOB_CLIP_EPSILON=0.2
-PPOB_ENTROPY_COEFF=0.001
+PPOB_ENTROPY_COEFF=0.0
 PPOB_VALUE_LOSS_COEFF=0.5
 PPOB_MAX_GRAD_NORM=0.5
 PPOB_GAMMA=0.99
@@ -12,7 +12,7 @@ PPOB_LAM=0.95
 PPOB_EPOCHS=4
 PPOB_MINI_BATCH_SIZE=64
 PPOB_LOG_STD_FLOOR=-3.0
-PPOB_LOG_STD_CEILING=-0.5
+PPOB_LOG_STD_CEILING=-1.0
 PPOB_INITIAL_LOG_STD=-0.5
 
 # PPOB_EVAL_ONLY=1 → freeze training (pure evaluation)

tools/warm_start.py

@@ -37,10 +37,10 @@ for name in unchanged:
     np.save(DST / f"{name}.npy", data)
     print(f"  {name}: {data.shape} copied")
 
-# Initialize log_std to -1.0 (std ≈ 0.37) — snapshot values (2.27–4.68) are too high for fine-tuning
+# Initialize log_std to -2.0 (std ≈ 0.135) — tighter than -1.0, proven workable
-log_std = np.full(6, -1.0, dtype=np.float32)
+log_std = np.full(6, -2.0, dtype=np.float32)
 np.save(DST / "log_std.npy", log_std)
-print(f"  log_std: initialized to -1.0 (std≈0.37), shape={log_std.shape}")
+print(f"  log_std: initialized to -2.0 (std≈0.135), shape={log_std.shape}")
 
 # Copy unchanged Adam moments (all except w1, which were handled above)
 unchanged_adam = [