A Reproducible Protocol for Benchmarking Inference Efficiency of Sub-3B Open-Weight Language Models on Commodity GPUs, with Reference Estimates | Synapse