Accurate quantification of hydrological fluxes in arid ecosystems is fundamental for modeling regional water cycle dynamics under progressive global aridification. Modeling these fluxes is complicated by the 'evaporation paradox,' a hydro-meteorological anomaly where actual evapotranspiration (ET) stagnates or declines despite rising atmospheric evaporative demand. Existing land surface models often lack the capacity to resolve the complex, nonlinear interactions between sparse vegetation canopies and soil water deficits. To address these structural limitations, we propose a Physiologically Constrained Shuttleworth-Wallace (PC-SW) two-source model tailored for temperate desert shrublands (Haloxylon ammodendron). The model integrates three physical modifications: (1) A piecewise linear soil resistance function was formulated to account for the disruption of liquid water continuity in shallow soil layers during desiccation, capturing the abrupt 'two-stage switching' effect triggered by dry surface layer (DSL) formation. (2) A hydraulic-stomatal regulation module, incorporating deep root zone moisture signals and Leaf Area Index (LAI) thresholds, was established to quantitatively characterize plant "physiological dormancy" under extreme drought. (3) A clumping index correction based on geometric optics was applied to rectify radiative transfer errors associated with patchy shrub canopies. Model performance was evaluated using high-frequency eddy covariance and sap flow data collected across two growing seasons (2023-2024) in the Minqin Oasis-Desert Transition Zone, China. Results indicate that the PC-SW model mitigates the systematic overestimation of soil evaporation inherent in the standard S-W model (reducing bias by similar to 45%) and achieves high simulation accuracy (R-2 > 0.90). Hydrological flux partitioning indicated that plant transpiration serves as the primary pathway for ecosystem water loss-contributing over 96% during the peak growing season-and is sustained largely by deep soil water reservoirs (40-300 cm). This analysis explicitly identifies the localized thermodynamic feedbacks driving the evaporation paradox. While the concept of "functional decoupling" induced by high surface resistance is recognized in ecohydrology, this study advances its predictive modeling by mathematically formalizing the specific physical and physiological thresholds that trigger this process. By incorporating a piecewise linear soil resistance model and an explicit dormancy threshold, the PC-SW model better captures the abrupt restriction of actual ET under severe water limitation. Consequently, excess radiant energy is preferentially partitioned into sensible heat, providing a mechanistic, threshold-driven explanation for the local evaporation paradox.