Accurate point spread function (PSF) calibration is important for camera performance evaluation, optical characterization, and physically grounded image restoration. The conventional slanted-edge method provides repeatable high-contrast transitions and dense sub-pixel sampling, but a single edge mainly characterizes the response along one normal direction and provides limited constraints for recovering anisotropic local PSFs. In this work, we propose CircleFlow, a framework for spatially varying camera PSF calibration based on omnidirectional slanted-edge modeling with a circle grid target. Each circular boundary is treated as a collection of local slanted-edge observations whose normal directions vary continuously along the contour. CircleFlow constructs a binary structural proxy from each captured pattern and refines its boundary geometry using dense optical flow. An energy-constrained implicit neural representation is employed to aggregate the orientation-varying edge responses into a physically valid PSF. Boundary refinement and PSF estimation are jointly optimized within an analysis-by-synthesis imaging model incorporating convolution, Bayer sampling, and demosaicing. Experiments on simulated and real camera data demonstrate that CircleFlow recovers complex and spatially varying PSFs more accurately than representative target-based methods. The frequency responses derived from the estimated PSFs agree closely with standard slanted-edge SFR measurements. Cross-device experiments further demonstrate the applicability of the recovered PSFs to full-field optical characterization and image restoration.