Beacon: Single-Turn Diagnosis and Mitigation of Latent Sycophancy in Large Language Models