Anomaly detection in autonomous vehicle (AV) telemetry is a safety-critical task requiring accurate identification of abnormal vehicle behavior from multivariate kinematic sensor streams. This study presents a comprehensive comparative benchmark of twelve supervised classification algorithms and thirteen unsupervised clustering configurations applied to a real-world AV telemetry dataset comprising 112,028 observations and 23 engineered kinematic features after removal of seven constant-value columns. The dataset exhibits an approximately balanced class distribution (Normal: 56,380; Anomaly: 55,648). All classifiers were trained on an 80/20 stratified split and evaluated on Accuracy, Precision, Recall, F1-Score, and AUC-ROC. Clustering algorithms were evaluated on 15,000 randomly sampled observations using Silhouette Score, Davies-Bouldin Index (DBI), Calinski-Harabasz Index (CHI), Adjusted Rand Index (ARI), and Normalized Mutual Information (NMI). Results demonstrate that tree-based ensemble classifiers achieve performance approaching theoretical upper bounds — Extra Trees, Random Forest, Bagging, and Decision Tree (each F1 = 0.9998) — driven by the dominant discriminative power of speed, speed_time_ratio, rolling_distance_mean, and distance, which collectively account for 59.7% of Random Forest feature importance. In contrast, all unsupervised clustering algorithms fail to semantically recover ground-truth anomaly labels: the best-Silhouette configurations (Birch k=3, Silhouette = 0.9529) achieve ARI ≈ 0.000, while the best-ARI configuration (K-Means k=2, ARI = 0.440) recovers only partial anomaly structure. Three robustness experiments qualify these figures: 5-fold cross-validation confirms stability under random partitioning (F1 = 0.9998 ± 0.0001 for the top ensembles), while strictly temporal and unseen-scenario partitioning reduce peak F1 to 0.9750 and 0.9602 respectively, quantifying the contribution of window- and scenario-level information leakage to the near-saturated hold-out scores; a ten-sample sensitivity analysis shows the K-Means label alignment to be sample-dependent (ARI = 0.20 ± 0.21) whereas the degenerate high-Silhouette collapse is fully robust. Under the default configurations and static distribution assumption examined here, these findings indicate that supervised classification is the preferred paradigm for AV anomaly detection when labeled telemetry is available, and establish that Silhouette Score alone is a misleading model selection criterion in this context. All experiments were run with default hyperparameters in a documented software environment; robustness under hyperparameter tuning and distribution drift remains for future work.