Sight-singing is a fundamental yet challenging task in music education, where timely and fine-grained feedback is crucial for helping learners identify mistakes and improve their performance efficiently. However, existing automatic methods mainly focus on isolated subtasks such as alignment or scoring, and are unable to provide end-to-end diagnostic feedback with explicit identification of pitch, rhythm, and musical fluency errors. In this paper, we propose Cantus, a novel multimodal large language model framework for sight-singing assessment and exercise recommendation generation. Cantus jointly models vocal audio and symbolic musical representations, and leverages a pedagogically grounded error representation module together with a large language model to diagnose performance errors, generate structured assessment comments, and provide personalized exercise recommendations. To support this task, we construct a new dataset MIVO (MIDI-Voice) containing 1,003 real-world sight-singing samples with aligned MIDI notes and vocal recordings collected from authentic teaching scenarios. Experimental results demonstrate the effectiveness of Cantus in sight-singing assessment, achieving strong performance across multiple metrics. Human evaluation of feedback quality shows the potential of Cantus in supporting student-centered sight-singing learning.