diff --git a/crates/specialized/rtx-cfd/src/kernels/cuda/reduction.cu b/crates/specialized/rtx-cfd/src/kernels/cuda/reduction.cu
index 204077e..a3c20cf 100644
--- a/crates/specialized/rtx-cfd/src/kernels/cuda/reduction.cu
+++ b/crates/specialized/rtx-cfd/src/kernels/cuda/reduction.cu
@@ -9,7 +9,7 @@
  * Using warp shuffle operations for modern GPUs
  */
 
-#include <cuda_runtime.h>
+// (no host headers: NVRTC compiles these as device code with the built-ins only)
 #include <float.h>
 
 // Warp size constant