Add debug OOB check for sorted_token_ids

vllm/nvfp4_cutedsl.py

@@ -340,6 +340,9 @@ class CuTeDSLMoERunner:
         expert_offsets[1:self.num_experts + 1] = tokens_per_expert.cumsum(0)
         
         # -- Gather hidden states into slot order --
+        # DEBUG: Check for OOB indices before indexing
+        if sorted_token_ids.max() >= num_tokens or sorted_token_ids.min() < 0:
+            print(f"[CLAWMINE BUG] sorted_token_ids OOB! max={sorted_token_ids.max().item()} min={sorted_token_ids.min().item()} num_tokens={num_tokens} _token_indices_max={self._token_indices[:num_slots].max().item()}", flush=True)
         slot_hidden = hidden_states[sorted_token_ids]
         
         # === L1: gate + up ===