Post-hoc model-agnostic explainers are widely used to interpret black-box machine learning models, yet practitioners often face contradictory attributions for the same prediction, while prior comparative studies yield inconsistent recommendations. We address this problem through a systematic comparison of seven model-agnostic explainers (LIME, Kernel~SHAP, Permutation~Importance, Partial~Dependence, Accumulated~Local~Effects, Occlusion, Anchors) and three gradient-based methods (Integrated~Gradients, Grad-CAM, SmoothGrad) across five tabular datasets, two image datasets, six models, and five random seeds, producing 760 experimental cells. Fidelity is evaluated using eight complementary metrics spanning perturbation faithfulness, local-surrogate fit, infidelity, stability, and ground-truth feature recovery on synthetic data, with statistical significance assessed through Friedman omnibus tests and Nemenyi post-hoc Critical Difference analysis. Three findings emerge. First, inter-explainer agreement depends on what is measured: agreement on which features matter is uniformly positive across all seven tabular explainers (+0.36 to +0.82), with no outlier. Agreement on direction of influence is interpretable only among explainers that produce signed attributions; there, Kernel~SHAP and Occlusion agree (+0.70), while LIME tracks neither (-0.06, -0.03), making LIME magnitude-consistent but sign-inconsistent with other local methods. Second, fidelity metrics can reverse rankings: Occlusion and Kernel~SHAP dominate perturbation-based metrics, whereas Anchors dominates ground-truth feature recovery, with Occlusion ranking last. Third, perturbation-fidelity rankings are baseline-dependent, making Kernel~SHAP the more baseline-robust default. We synthesize these findings into a conditional decision guide and release the full reproducible pipeline.