fix(zerogpu): add explicit duration, disable geometry offload by default

ZeroGPU default lease is ~60s. Hunyuan3D .to("cuda") (~10-15 GB H2D
transfer) + 100-step diffusion easily exceeds that, causing the callback
to be killed silently at ~10/100 steps.

- Add _gpu(duration) helper; replace @GPU with @_gpu(duration=N) for each
callback: generate_mesh/generate_slat 240s, render_* 180s, videos 300s.
- Change NEAR_GEOMETRY_OFFLOAD_AFTER_MESH default to "0": H200 has 80 GB
VRAM; keeping Hunyuan on GPU after mesh generation means subsequent
calls only pay a no-op .to("cuda") instead of a full ~30s H2D transfer.
Override with NEAR_GEOMETRY_OFFLOAD_AFTER_MESH=1 if needed.

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

app.py

@@ -59,6 +59,13 @@ from trellis.pipelines import NeARImageToRelightable3DPipeline
 
 GPU = spaces.GPU if spaces is not None else (lambda f: f)
 
+
+def _gpu(duration: int = 120):
+    """ZeroGPU decorator with explicit duration; no-op when spaces is absent."""
+    if spaces is not None:
+        return spaces.GPU(duration=duration)
+    return lambda f: f
+
 APP_DIR = Path(__file__).resolve().parent
 CACHE_DIR = APP_DIR / "tmp_gradio"
 CACHE_DIR.mkdir(exist_ok=True)
@@ -84,9 +91,7 @@ def _truthy_env(name: str, default: str) -> bool:
 _CPU_PRELOAD_AT_START = _truthy_env(
     "NEAR_MODEL_CPU_PRELOAD_AT_START", "1" if spaces is not None else "0"
 )
-_OFFLOAD_GEOMETRY_AFTER_MESH = _truthy_env(
-    "NEAR_GEOMETRY_OFFLOAD_AFTER_MESH", "1" if spaces is not None else "0"
-)
+_OFFLOAD_GEOMETRY_AFTER_MESH = _truthy_env("NEAR_GEOMETRY_OFFLOAD_AFTER_MESH", "0")
 
 
 def _path_is_git_lfs_pointer(p: Path) -> bool:
@@ -286,7 +291,7 @@ def _require_hdri_path(hdri_obj: Any) -> str:
     return p
 
 
-@GPU
+@_gpu(duration=240)
 @torch.inference_mode()
 def generate_mesh(
     image_input: Optional[Image.Image],
@@ -326,7 +331,7 @@ def generate_mesh(
     return state, str(mesh_path), "**① Mesh ready** — run **② SLaT** next."
 
 
-@GPU
+@_gpu(duration=240)
 @torch.inference_mode()
 def generate_slat(
     asset_state: Dict[str, Any],
@@ -390,7 +395,7 @@ def load_slat_npz(
     return state, f"SLaT loaded: **{Path(resolved).name}**"
 
 
-@GPU
+@_gpu(duration=180)
 @torch.inference_mode()
 def render_preview(
     asset_state: Dict[str, Any],
@@ -434,7 +439,7 @@ def render_preview(
     )
 
 
-@GPU
+@_gpu(duration=180)
 def export_pbr_glb(
     asset_state: Dict[str, Any],
     hdri_file_obj: Any,
@@ -471,7 +476,7 @@ def export_pbr_glb(
     return str(out), f"**③ PBR GLB** — `{out.name}`"
 
 
-@GPU
+@_gpu(duration=300)
 @torch.inference_mode()
 def render_dual_lighting_videos(
     asset_state: Dict[str, Any],