Accurately detecting student affective states is critical for building adaptive e-learning systems. However, the field faces a replicability crisis as results are often highly sensitive to overlooked hyperparameter choices. This study presents a systematic sensitivity analysis to isolate the co-dependency of three critical factors: (1) feature sets (Landmarks, FAUs, and Combined), (2) feature representations (Static vs. Dynamic deltas), and (3) label binarization thresholds. To ensure a controlled environment for measuring these variables, a rigid Start Middle-End (SME) temporal sampling strategy was employed as a baseline. Models were trained on the DAISEE dataset using SMOTE and subject-aware validation to address severe class imbalance and generalizability. Our findings reveal that the ‘optimal’ configuration is highly volatile and unique for each affective state, reinforcing a ‘model-per-emotion’ requirement. While Boredom achieved a 0.63 F1-score with static features, Engagement required a dynamic delta-Cosine representation (0.55 F1) on a balanced threshold. By exposing how arbitrary label definitions and feature metrics can dramatically alter performance, this work provides a methodological roadmap for building more robust persuasive technologies.