Existing deep learning methods for remote photoplethysmography (rPPG) struggle with robust signal extraction due to a lack of explicit constraints for noise disentanglement. We propose a novel hybrid architecture that introduces a physics-informed reconstruction module to address this limitation. This module guides the network to decompose the signal by reconstructing the raw measurement traces using a fixed pulse and learnable motion basis vector. The architecture also integrates a temporal differencing front-end and is optimized with a hybrid time-frequency loss. Our model achieve superior performance on the PURE and UBFC-rPPG datasets, significantly improving SNR by over 2.5 dB on PURE, and secures top performance on UBFC-rPPG by over 0.26 dB.