We propose a temporal disambiguation physicallayer authentication framework for non-stationary multiple-input multiple-output channels, where temporal correlation and longterm line-of-sight blockage violate the independence, identicaldistribution, and stationarity assumptions of classical physicallayer authentication. Our key idea is to resolve the ambiguity between channel-induced and adversarial distribution shifts via a context-aware, posterior-driven inference mechanism. Specifically, we adopt a state-augmented sequential formulation in which line-of-sight availability is modeled as a latent channel state and jointly inferred with adversarial activity, while posteriorweighted exponential moving averages continuously update the emission statistics to track non-stationary channel drifts. To stress-test robustness against adaptive attackers, we consider distribution-matching spoofers based on generative models, potentially with temporal correlation to mimic channel dynamics. Closed-form recursions for the resulting decision statistic and tractable operating-characteristic approximations are derived, showing improved robustness over sequential independent and identically distributed authentication baselines.