IDU 0.2.0: add decider-4b as a second backend

app.py

@@ -749,17 +749,18 @@ def health():
                      "the GPU worker and model_in_web_process will read False by design.")}
 
 
+@GPU
 def warmup_decider() -> dict:
     """Build decider-4b and run one real decision, so both backends are warm at startup.
 
-    This is the part that costs real memory: laya is 421M and decider is 4.2B, and both sit
-    in the same Space. Kept separate from `warmup` so a decider failure cannot take Laya
-    down with it -- the two are independent services sharing a UI.
+    DECORATED WITH @GPU, which is not optional here. Without it this runs outside any GPU
+    context: on ZeroGPU `torch.cuda.is_available()` can still report True while no device is
+    actually attached, so moving the model fails inside the CUDA runtime with
+    "no CUDA-capable device is detected". The decorated call is the only place a device
+    exists.
 
-    The model is built on CPU at module scope and moved onto the GPU HERE, which mirrors the
-    Laya path: on ZeroGPU a device exists only inside a decorated call, so building on
-    "cuda" at import time is not possible. `_decider_move_to_gpu` degrades to CPU rather
-    than raising, so this returns a result either way and the device is reported.
+    Kept separate from `warmup` so a decider failure cannot take Laya down with it -- the
+    two are independent services sharing a UI.
     """
     try:
         get_decider()