Mechanism-informed models that can simulate counterfactual mental-health trajectories remain scarce in digital phenotyping. Most existing approaches either predict outcomes from sensor-derived data without specifying the cross-domain generative process or reconstruct latent dynamics without encoding the mechanistic assumptions needed for intervention simulation. Here, we introduce the Coupled Stochastic Dynamical System (CSDS), a forward generative model that jointly simulates digital engagement, weara-ble physiology, and psychiatric burden as a family of coupled stochastic processes. Using the multi-year GLOBEM cohort (N=496) integrating passive sensing, ecological momentary assessment, and survey-based measures, we first identified four distinct behavioural phenotypes by clustering. Bayesian inversion of the CSDS satisfied 80-83% of clusters' weighted phenotype targets through separable trait-vulnerability, physical-activity, and affective-response parameter families. Participant-level inversion produced virtual twins that recovered the individual dynamic structure of complex human behaviours: Hidden Markov Models trained on the synthetic timeseries decoded observed state occupancy and transi-tions (r=0.84-0.87 and 0.93-0.94), and outperformed mean- or persistence-based null models in short-horizon magnitude prediction, with sparser affective channels showing horizon-stable directional forecast-ing above chance (66-78% balanced accuracy to 21 days). These findings position generative behaviour-al models as falsifiable and interpretable frameworks for studying youth mental-health trajectories and for developing future personalised interventions.