fix(SAC_LSTM_Bot): checkpoint save check inside gradient-step loop (#56)

Launch finding during campaign-v1 verification: at production sizes
(hidden 256, ~1s/step, ~13s trainer CPU per ~40s chunk process) the
save check ran only between drain-burst passes, so stepCount never
crossed nextSave before the process died — zero checkpoints persisted
across entire runs (masked at #49/#54 smoke sizes where steps were
sub-millisecond). Check now fires mid-loop; with SAVE_INTERVAL<=5
(within the per-process step budget) every chunk persists its chain.

SAC_LSTM_Bot/src/SAC_LSTM_Bot/integration.nim

@@ -290,6 +290,15 @@ proc trainPass*(st: var TrainState; drained: int) =
       break
     discard sacUpdate(st.trainer, seqs)
     inc st.stepCount
+    # Save check INSIDE the step loop (#56 launch finding): at production sizes
+    # (hidden 256 ⇒ ~1 s/step) a drain burst queues minutes of steps; checking
+    # only between passes meant the process died mid-loop before stepCount ever
+    # reached nextSave — zero checkpoints persisted for the whole campaign.
+    # Mid-loop checks + SAVE_INTERVAL≤20 (#54) keep saves ~20 s apart.
+    if st.stepCount >= st.nextSave:
+      st.nextSave += getSaveInterval()
+      var full = packFull(st.trainer)
+      discard gSaveChan.trySend(move(full))   # cap-1: drop if I/O thread is busy (Q5)
   # Publish latest actor (Q7): in-place write under the lock, bump version.
   withLock(gWeightLock):
     assert gSharedSnap.hiddenDim == st.trainer.actor.hiddenDim,
@@ -297,10 +306,6 @@ proc trainPass*(st: var TrainState; drained: int) =
     var c = 0
     packActor(st.trainer.actor, gSharedSnap.data, c)
     inc gSharedSnap.version
-  if st.stepCount >= st.nextSave:
-    st.nextSave += getSaveInterval()
-    var full = packFull(st.trainer)
-    discard gSaveChan.trySend(move(full))   # cap-1: drop if I/O thread is busy (Q5)
 
 # ── Threads ───────────────────────────────────────────────────────────────────