Reliability is required for crack path prediction under fretting fatigue in safety critical applications, yet it is often difficult to achieve. Here, we propose Trustworthy Inference Diffusion(TriDiff), a kinematics-aware generative AI framework that transitions from deterministic prediction to probabilistic physical reasoning. Unlike standard vision-based models that rely solely on geometric features, the initial displacement and stress fields are used as conditioning priors. Our methodology employs a stochastic sampling protocol to obtain the conditional probability distribution of crack propagation rather than a single path. First, a finite element model is validated through comparison with experimental observation. Then, to evaluate accuracy, TriDiff is benchmarked against five baseline models, including ConvLSTM, PredRNN, DeepLabv3, SegNet, and UNet. Crack tip error is reduced by more than a factor of three when the best baseline is used as the reference. The mean absolute error of crack length is also reduced by more than 2.6 times relative to the best baseline. Third, by visualizing epistemic uncertainty, we show that the model's variance is not random noise but forms a coherent confidence map. Crucially, high variance regions are localized precisely within the fracture process zone, which indicates that regions of physical instability are implicitly identified by the model. Finally, quantitative analysis reveals a fail safe property: even when the deterministic mean prediction deviates from the ground truth in complex scenarios, the true crack path is contained within the model's 95% confidence interval with a coverage rate above 98%.