feat(PPO_Bot): deterministic eval + fix logStd warm-start

- actorForward: deterministic param, uses mean-only when PPOB_EVAL_ONLY=1
  (eval was adding unit Gaussian noise to every action — unreliable scores)
- warm_start.py: log_std initialized to -1.0 (std≈0.37) instead of copying
  snapshot values (were 2.27-4.68 → std 9-108, completely drowning signal)
- training.env: LOG_STD_CEILING 0.0→-0.5 (cap exploration at std≈0.6)

PPO_Bot/PPO_Bot.nim

@@ -281,7 +281,7 @@ method run(bot: PPOBot) =
     # detection-tick pulse, and the next iteration's spawn check sees false —
     # one shot → exactly one bullet, even across deadReckon gaps.
     bot.tracker.current.hasFired = false
-    let (rawActs, logP) = ac.actorForward(state)
+    let (rawActs, logP) = ac.actorForward(state, deterministic = evalOnly)
     let value = ac.criticForward(state)
     let ex = if bot.tracker.hasContact: bot.tracker.current.x else: botData.arenaWidth  / 2.0
     let ey = if bot.tracker.hasContact: bot.tracker.current.y else: botData.arenaHeight / 2.0

PPO_Bot/network.nim

@@ -43,9 +43,13 @@ proc forward*(mlp: MLP, x: Tensor[float32]): Tensor[float32] =
   let h2 = tanh(mlp.w2 * h1 + mlp.b2)
   result  = mlp.w3 * h2 + mlp.b3
 
-proc actorForward*(ac: ActorCritic, state: Tensor[float32]): tuple[actions: Tensor[float32], logProb: float32] =
+proc actorForward*(ac: ActorCritic, state: Tensor[float32], deterministic = false): tuple[actions: Tensor[float32], logProb: float32] =
-  ## state: [STATE_DIM]. Returns sampled actions [ACTION_DIM] and sum log-prob.
+  ## state: [STATE_DIM]. Returns actions [ACTION_DIM] and sum log-prob.
+  ## deterministic=true: return mean only (no noise), logProb=0.
   let mean = ac.actor.forward(state)
+  if deterministic:
+    return (actions: mean, logProb: 0.0'f32)
+
   # Floor logStd at logStdFloor before exp → min std ≈ exp(logStdFloor)
   var clampedLogStd = ac.logStd.map(proc(v: float32): float32 = clamp(v, logStdFloor, logStdCeiling))
   let std = clampedLogStd.map(proc(v: float32): float32 = exp(v))

tools/training_runner/training.env

@@ -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=0.0
+PPOB_LOG_STD_CEILING=-0.5
 PPOB_INITIAL_LOG_STD=-0.5
 
 # PPOB_EVAL_ONLY=1 → freeze training (pure evaluation)

tools/warm_start.py

@@ -31,13 +31,17 @@ for prefix in ("actor", "critic"):
 unchanged = [
     "actor_w2", "actor_w3", "actor_b1", "actor_b2", "actor_b3",
     "critic_w2", "critic_w3", "critic_b1", "critic_b2", "critic_b3",
-    "log_std",
 ]
 for name in unchanged:
     data = np.load(SRC / f"{name}.npy")
     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
+log_std = np.full(6, -1.0, 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}")
+
 # Copy unchanged Adam moments (all except w1, which were handled above)
 unchanged_adam = [
     "adam_aw2", "adam_cw2",