fix(PPO_Bot): SIGSEGV crash fixes + static buffers for thread safety

- bullets: seq[InFlightBullet] → array[4, InFlightBullet] + bulletCount
  (eliminates cross-thread heap realloc under ORC)
- hasFired: edge-triggered (cleared after state build, not level-triggered)
- round_counter parseInt: wrapped for empty/torn file → 0
- Static SVG + intent buffers to kill cross-thread heap realloc
- Tick-local alive/bulletData also fixed arrays

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

PPO_Bot/tests/test_training.nim

@@ -149,4 +149,27 @@ block testPpoUpdateNormalReward:
   check m.valueLoss == m.valueLoss, "valueLoss NaN on normal-reward round"
   check m.gradNorm  == m.gradNorm,  "gradNorm NaN on normal-reward round"
 
+# ── logStd ceiling: raw param must never drift above the collection clamp ─────
+# Regression for the train/collection std mismatch: logStd starting above the
+# ceiling (as in the warm-start snapshot) must be clamped back to [floor, ceiling]
+# by the first Adam step, so recomputed logP matches the acting policy's std.
+
+block testLogStdCeilingClamp:
+  randomize(45)
+  var ac = initActorCritic()
+  ac.logStd = newTensor[float32](ACTION_DIM).map(proc(v: float32): float32 = 3.0'f32)
+  var buf = initTrajectoryBuffer()
+  for _ in 0..<16:
+    let s = randomNormalTensor[float32](STATE_DIM)
+    let a = randomNormalTensor[float32](ACTION_DIM)
+    let lp = ac.computeLogProb(s, a)
+    buf.add(Transition(state: s.stateToArr, action: a.actionToArr, logProb: lp,
+                       reward: 0.1'f32, value: 0.5'f32))
+  var adam: ACAdamStates
+  discard ppoUpdate(ac, buf, lastValue = 0.0'f32, adamStates = adam,
+                    epochs = 1, miniBatchSize = 16)
+  for v in ac.logStd:
+    check v <= logStdCeiling + 1e-6'f32, "logStd above ceiling after ppoUpdate"
+    check v >= logStdFloor - 1e-6'f32, "logStd below floor after ppoUpdate"
+
 echo "All tests passed"