We present a lightweight and computationally efficient system for Subtask 1 of SemEval-2026 Task 2, which focuses on predicting longitudinal variation in emotional valence and arousal from ecological essays. Our approach uses frozen contextual embeddings from BERT-base-uncased to obtain mean-pooled sentence representations without fine-tuning the transformer. These 768-dimensional embeddings are fed into a multi-output Ridge regression model to jointly predict normalized valence and arousal scores.The system emphasizes simplicity, reproducibility, and efficiency, avoiding complex temporal architectures, external lexicons, or user metadata. Despite its simplicity, the model achieves strong performance for valence prediction (r = 0.594) and moderate performance for arousal prediction (r = 0.296). Detailed evaluation across seen and unseen users, as well as between-user and within-user splits, shows that between-user correlations are consistently higher, and that valence is substantially easier to predict than arousal. These findings suggest that frozen transformer embeddings combined with linear regression provide a competitive and interpretable baseline for longitudinal affect prediction tasks.