To handle nonstationary dynamics and intersubject variability in fNIRS depression classification, we fuse temporal, spectral, spatial, and cross-channel features and design a multiscale DTW with temporal and channel attention. Individual and cohort templates are used jointly to capture fine-grained descriptors and subject-specific deviations. Feature subsets are selected using TopCombo, and classification employs a stacking ensemble with SVM, Random Forest, and XGBoost as base learners, along with a logistic regression meta-learner. Evaluation follows a strict subject-wise protocol to avoid leakage. All transformations and hyperparameters are fit within training folds, and inference is performed only on held-out subjects. On a cohort of 56 MDD patients and 53 healthy controls, the proposed framework achieves an accuracy of 0.90 and a macro F1-score of 0.898 under strict subject-wise cross-validation. Compared with representative baselines, macro F1 improves by about 4.3–7.5 percentage points, and the variance across folds is clearly reduced. These results indicate that coupling multi-view features with personalized multiscale alignment is critical for robust cross-subject generalization in fNIRS-based depression recognition.