feat(PPO_Bot): weight persistence + background training (#17)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

PPO_Bot/training.nim

@@ -149,9 +149,11 @@ proc initACAdamStates(ac: ActorCritic): ACAdamStates =
   result.cb3    = initAdamState(ac.critic.b3)
   result.logStd = initAdamState(ac.logStd)
 
-# Persistent Adam state — survives across ppoUpdate calls (lives in training module)
-var gAdamStates: ACAdamStates
-var gAdamInit = false
+# Persistent Adam state — per-thread so training thread can call ppoUpdate safely.
+# ponytail: threadvar resets Adam each new training thread; persist across calls
+# within one thread. If Adam across rounds matters, embed state in TrainingArgs.
+var gAdamStates {.threadvar.}: ACAdamStates
+var gAdamInit   {.threadvar.}: bool
 
 # ── Gradient clipping ─────────────────────────────────────────────────────────
 
@@ -172,7 +174,7 @@ proc ppoUpdate*(ac: var ActorCritic;
                 entropyCoeff: float32 = 0.01'f32;
                 valueLossCoeff: float32 = 0.5'f32;
                 lr: float32 = 3e-4'f32;
-                maxGradNorm: float32 = 0.5'f32) =
+                maxGradNorm: float32 = 0.5'f32) {.gcsafe.} =
   if buffer.len == 0: return
 
   # Initialise Adam states once (persists across rounds)