Mobile health (mHealth) recommender systems for physical activity increasingly employ sophisticated personalization, yet few integrate clinical-guideline-grounded user modeling, explainability, and fairness auditing end-to-end, and fewer still evaluate how users perceive such systems in real community settings. This paper presents FitWisdom, a personalized exercise recommender whose user model maps demographic and clinical attributes to authoritative guidelines (WHO, ACSM, AHA, NKF, ADA) and trains a contextual multi-armed bandit (LinUCB) over Deep-Q-Network-generated synthetic profiles for cold-start. Recommendations are accompanied by SHAP-based natural-language rationales and audited for demographic fairness via the Equal Opportunity Difference metric. We evaluated user perceptions with 32 adults at a YMCA facility using a mixed-methods design combining the Health Information Technology Usability Evaluation Scale (Health-ITUES) framework with Partial Least Squares Structural Equation Modeling (PLS-SEM, 1,000 bootstrap resamples) and concurrent think-aloud sessions. The structural model explained 71.4% of variance in Perceived Usefulness, with Impact mediating the effects of Perceived Ease of Use (β = 0.460, p = 0.006) and User Control (β = 0.323, p = 0.030). Qualitative analysis identified onboarding friction and feature discoverability as the primary perception-shaping mechanisms, issues with direct implications for user-model fidelity in a clinically grounded system. We position these findings as complementary to recent longitudinal evaluations of recommender algorithms for physical activity in this journal, together sketching a more complete evaluation portfolio for mHealth recommenders.