Modeling longitudinal affect requires capturing both stable user tendencies and transient textual signals. For SemEval-2026 Task 2, we propose the Trait-State Affect Forecaster (TSAF), which decomposes affect into persistent user traits and text-conditioned states integrated through adaptive gating. On per-text prediction (Subtask 1), TSAF achieves composite Pearson correlations of 0.645 for valence and 0.409 for arousal, outperforming the Linear(BERT) baseline. In forecasting tasks, results reveal strong short-term affective inertia, where prior affect dominates next-step prediction, while long-term drift remains challenging under sparse supervision; TSAF shows comparatively stronger gains for arousal in this setting. Analyses across user splits and modalities highlight the strengths and trade-offs of explicit trait-state modeling, particularly under cold-start and short-text conditions.