<p>The validity of functional linear models, which link a scalar response to a functional predictor through a linear integral operator, relies critically on two core assumptions: the independence of the predictor from the error term and the linearity of the regression operator. However, due to their inherent interdependence, separately testing these assumptions presents a significant challenge. In this paper, we propose a new framework for jointly testing the independence of predictors and errors and the goodness-of-fit in functional linear models, based on a new dependence measure termed Rank-Kernel Divergence (RKD). By combining rank- and kernel-based approaches, the resulting RKD measure is well suited to scalar-response problems and possesses several attractive properties: it requires no moment restrictions, is invariant under monotone transformations, and leads to consistent tests. We construct a test statistic based on RKD, study its asymptotic properties, and evaluate its finite-sample performance via simulations. Numerical experiments show that the proposed test is competitive across a variety of scenarios and often outperforms existing methods. The practical utility of the method is illustrated via two real-data analyses.</p>