Calibrating hybrid and parallel micro-nano positioning robots requires measured pose errors to be transferred to identifiable geometric quantities without violating closed-loop compatibility or the equality of components shared among kinematic paths. Existing projection and rank-reduction methods do not generally distinguish physical-sharing constraints, unmeasured passive-state directions, and measurement-null parameter directions within one model. This paper develops a metrology-assisted, constraint-preserving framework based on local product-of-exponentials kinematics represented by unit dual quaternions (UDQs). A digital holographic microscope and a pseudo-periodic encoded target provide local six-degree-of-freedom pose measurements. Binary assembly maps first assign one global variable to each physical component, and a limb-incidence operator expresses loop compatibility. A full-SVD left-null-space projection then eliminates passive-coordinate perturbations without constraint-matrix inversion; a second SVD constructs rank-minimal observable coordinates within the retained physical-sharing manifold. The same projected feasible space defines a damped active-joint update and separates compensable from irreducible task errors. On a 6-DOF parallel and a 5-DOF hybrid platform, 100 measured poses per platform were divided into 50 identification and 50 held-out poses within one continuous out-of-plane window. Mean held-out position/orientation residuals were 1.25 μm/28.3 μrad and 17.1 μm/354 μrad, respectively. Relative to an algebraically matched homogeneous-transformation-matrix implementation, UDQ reduced identification runtime by 23.1%-23.5%; a 1000-pose numerical timing study showed a 13.0%-20.0% reduction in compensation time. The evidence supports local cross-architecture predictive consistency and representation-level efficiency, but not independently traceable absolute or measured post-compensation accuracy.