This article introduces a new class of recursive linearly constrained minimum variance estimators (LCMVEs) that provides additional robustness to modeling errors. To achieve that robustness, a set of non-stationary linear constraints are added to the standard LCMVE that allow for a closed form solution that becomes appealing in sequential implementations of the estimator. Indeed, a key point of such recursive LCMVE is to be fully adaptive in the context of sequential estimation as it allows optional constraints addition that can be triggered by a preprocessing of each new observation or external information on the environment. This methodology has significance in the popular problem of linear regression among others. Particularly, this article considers the general class of partially coherent signal (PCS) sources, which encompasses the case of fully coherent signal (FCS) sources. The article derivates the recursive LCMVE for this type of problems and investigates, analytically and through simulations, its robustness against mismatches on linear discrete state-space models. Both errors on system matrices and noise statistics uncertainty are considered. An illustrative multi-channel array processing example is treated to support the discussion, where results in different model mismatched scenarios are provided with respect to the standard case with only FCS sources.