This paper presents a rigorous, mathematically grounded framework for deploying high-fidelity dermatological classification and biometric analysis directly on resource-constrained mobile hardware. Addressing the fundamental trilemma of data scarcity, computational latency, and privacy in mHealth applications, we introduce a novel hybrid learning pipeline. We formulate a robust approach utilizing Generative Adversarial Networks (GANs) with Adaptive Differentiable Augmentation (ADA) for synthetic data generation, explicitly addressing the long-tailed class imbalance inherent in medical datasets and the under-representation of diverse skin tones. Furthermore, we propose a latency-constrained Differentiable Neural Architecture Search (DNAS) utilizing a Lagrangian relaxation method to optimize network topology specifically for mobile inference. We provide an analytical derivation of quantization error bounds for INT8 mobile inference and introduce a Hessian-aware bit-width assignment strategy. Additionally, we model the thermodynamic behavior of mobile SoCs under continuous inference loads, deriving a control-theoretic approach to prevent thermal throttling. Experimental results on a composite dataset of 20,015 images demonstrate that our synthetic augmentation strategy improves mean Average Precision (mAP) by 14.3% in low-data regimes. Crucially, we present a demographic breakdown showing that our balanced synthetic training reduces the False Negative Rate (FNR) gap between Fitzpatrick Skin Types I-II and V-VI from 12.4% to 3.1%. Our architectural optimizations reduce inference latency by 45% (<12ms) on standard mobile DSPs compared to baseline EfficientNet architectures, enabling real-time, privacy-preserving diagnostics.