Transparent cells often exhibit weak and spatially heterogeneous directional-intensity responses in non-interferometric microscopy, making morphology reconstruction vulnerable to low signal, incomplete angular information, and boundary artifacts. We present a polarization-reliability-adaptive soft-constrained ring-illumination method (PRAS-RIM) for reconstructing the apparent relative 3D morphology of transparent samples. The principal acquisition configuration comprises a programmable 12-sector ring source (P12), a PET diffuser, a fixed circular polarizer, a 40× objective with NA = 0.65, and a division-of-focal-plane polarization camera. Calibrated directional intensity provides the primary geometric evidence, whereas the angle of polarization (AoP) is represented in double-angle form and incorporated only as a pixelwise reliability-weighted soft constraint. Across 29 microsphere measurements, complete PRAS-RIM reduced the mean object-level RMSE from 2.77 ± 0.19 μm for the intensity-only baseline to 0.59 ± 0.07 μm. A yeast test set comprising 51 cells from three distinct static fields acquired within the same experiment demonstrated cross-field consistency and spatially varying reliability, but was not used for absolute-height validation because no independent geometric ground truth was available. These results support PRAS-RIM as an interpretable framework for robust label-free reconstruction of relative morphology without claiming optical phase, optical path length, absolute cell height, or a three-dimensional refractive-index distribution.