Benchmarking the Explanatory Quality of Open-Weight Vision-Language Models in Face Recognition