Recognize api.cerebras.ai as a Cerebras cloud provider so llama.cpp/LM Studio cache-affinity fields are not attached even when endpoint_kind is misconfigured as local. Add regression coverage for provider detection, self-hosted classification, and payload field exclusion.