Auto-Generated CUDA Kernels Need Kernel-Level Validation

An LLM-written kernel benchmarked 38% faster on a microbench. Here is what kernel-level validation...