LLM inference gets zkSNARKed at operator granularity, because apparently GPUs weren't opaque enough.