node kit: alpha flag through train/eval, carry env+width fixes, bootstrap script

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

scripts/train_merge_code.py

@@ -40,9 +40,13 @@ ap.add_argument("--seed", type=int, default=0)
 ap.add_argument("--pause", type=int, default=0,
                 help="pause-token control: p inert tokens after prompt, "
                      "feedforward adapter, no recurrence")
+ap.add_argument("--alpha", type=float, default=0.3,
+                help="merge weight (2B-tuned default 0.3; try 0.1-0.15 at 12B)")
 ARGS = ap.parse_args()
 SEED = ARGS.seed
-SUFFIX = (f"_s{SEED}" if SEED else "") + (f"_p{ARGS.pause}" if ARGS.pause else "")
+SUFFIX = ((f"_s{SEED}" if SEED else "")
+          + (f"_p{ARGS.pause}" if ARGS.pause else "")
+          + (f"_a{ARGS.alpha}" if ARGS.alpha != 0.3 else ""))
 PAUSE_ID = 6  # <unused0>
 
 
@@ -101,7 +105,7 @@ def main():
         p.requires_grad_(False)
     looper = BandLooper(model)
     d = model.config.get_text_config().hidden_size
-    adapter = MergeAdapter(d=d).cuda()
+    adapter = MergeAdapter(d=d, alpha=ARGS.alpha).cuda()
     opt = torch.optim.AdamW(adapter.parameters(), lr=LR, weight_decay=0.01)
 
     train = [it for it in data if it["split"] == "train"