fix: surface bounded GPU admission failures

app.py

@@ -58,7 +58,7 @@ if "HF_HOME" not in os.environ:
 import gradio as gr
 import torch
 from obliteratus import device as dev
-from obliteratus.gpu_lifecycle import from_environment, measure_torch_memory
+from obliteratus.gpu_lifecycle import AdmissionError, from_environment, measure_torch_memory
 from obliteratus.benchmark_lifecycle import (
     admit_benchmark,
     mark_benchmark_ready,
@@ -2225,7 +2225,26 @@ def obliterate(model_choice: str, method_choice: str,
 
     # Admission must be granted before cleanup can touch the CUDA runtime or
     # the worker can enter any model-loading path.
-    _gpu_lifecycle.loading(model_id)
+    try:
+        _gpu_lifecycle.loading(model_id)
+    except AdmissionError as exc:
+        with _lock:
+            _state["status"] = "idle"
+            _state["model"] = None
+            _state["tokenizer"] = None
+            _state["log"] = [
+                f"GPU ADMISSION FAILED: {exc.user_message()}",
+                f"Diagnostic: {exc.diagnostic_message()}",
+            ]
+        yield (
+            f"**GPU admission failed:** {exc.user_message()}",
+            "\n".join(_state["log"]),
+            get_chat_header(),
+            gr.update(),
+            gr.update(),
+            gr.update(),
+        )
+        return
     _clear_gpu(release_lifecycle=False)
 
     with _lock: