Many deep unfolding networks for single-pixel imaging (SPI) are trained for a fixed compression ratio (CR). During the measurement process, block-based schemes suffer from blocking artifacts at low CRs, and standard Kronecker compressive sensing (KCS) does not specify a fixed-prefix acquisition sequence. To address these limitations, we propose NAR-Net, a physics-guided deep unfolding network trained once for reconstruction across arbitrary CRs. We first introduce nested Kronecker compressive sensing (NKCS), which pairs each row operator with a random-sign column modulation following an energy-ordered sequence. This construction guarantees that measurements at lower budgets form prefixes of those at higher budgets, thereby enabling the addition of new measurements without hardware reconfiguration. We then build a ratio-adaptive unfolding framework that injects the CR as a conditional scalar through adaptive layer normalization and separates local texture from global structure modeling via hybrid split blocks. Simulations and physical SPI experiments use a reconstruction size of 256 × 256. NAR-Net achieves a peak signal-to-noise ratio (PSNR) of 24.41 dB at CR = 0.01 on Set11, outperforming ProxUnroll by 1.65 dB. A single NAR-Net model covers continuous CRs ranging from 0.01 to 0.50 and matches or exceeds the performance of ratio-specific baselines at the majority of the evaluated CRs. Our codes and models are available at https://github.com/walker9921/NAR-Net.