Update Feather H200 training runtime image

overlay/scripts/run_domain_expanded_pretrain.sh

@@ -224,7 +224,13 @@ else
 fi
 log "note=train.py consumes HYDRA_RESUME_CKPT and HYDRA_CKPT_INTERVAL env vars; launcher exports them automatically"
 
-if [[ "$NEED_PREPARE" -eq 1 ]]; then
+if [[ "${HYDRA_USE_NEMOTRON:-0}" == "1" ]]; then
+  # Streaming Nemotron path (Super3 recipe) pulls tokens directly from HF at
+  # train-time via prepare_nemotron.make_dataloader. The disk-shard prepare.py
+  # download phase is redundant in this mode and wastes 20-30 min of paid GPU
+  # time on shard parquet transfers we'll never read.
+  log "prepare_action=skip reason=HYDRA_USE_NEMOTRON=1 (streaming at train-time)"
+elif [[ "$NEED_PREPARE" -eq 1 ]]; then
   PREPARE_CMD=("${PYTHON_CMD[@]}" prepare.py --num-shards "$PREPARE_NUM_SHARDS" --download-workers "$DOWNLOAD_WORKERS")
   log "prepare_action=run command=${PREPARE_CMD[*]}"
   if [[ "$DRY_RUN" -eq 0 ]]; then