Predicting the three-dimensional conformation of molecules from their bonding graph is a fundamental problem in computational chemistry, with applications in drug discovery, materials science, and molecular simulation. We present a bond-centric graph neural network framework for predicting local molecular geometry directly from graph connectivity. Instead of generating full Cartesian coordinates in a single step, the method decomposes the problem into local predictions around individual bonds, using a graph attention network to encode chemical environments and a rotationally invariant spherical representation to model bond distances and angular structure. To support this formulation, we construct an independently generated benchmark dataset of 9,691 DFT-optimized molecular structures using a large-scale Quantum ESPRESSO workflow initialized from PubChemQC molecular candidates. The resulting dataset provides rotationally consistent local geometric supervision for bond-centric molecular reconstruction while remaining independent from the original molecular optimization pipeline. On the benchmark set, the proposed framework achieves a heavy-atom bond-distance mean absolute error of 0.103 Å and a root mean squared error of 0.152 Å, corresponding to an average relative bond-length error of approximately 7.37%. Local angular reconstruction achieves a polar-angle θ mean absolute error of 4.56° and root mean squared error of 7.49°, while azimuthal-angle Δφ prediction remains substantially more challenging, with a mean absolute error of 18.19° and root mean squared error of 31.34°. Error analysis shows that bond lengths and local bond-angle magnitudes are reconstructed reliably across diverse chemical environments, whereas torsional and rotational degrees of freedom contribute the dominant source of geometric uncertainty, particularly in flexible single-bond conformations. Qualitative reconstruction results demonstrate strong preservation of molecular topology and local bonding environments, with successful extrapolation to polymeric systems significantly larger than those observed during training. From a computational perspective, the method exhibits predictable scaling with system size, following a power-law relationship of approximately T proportional to B^1.98 with respect to bond count. In direct comparison, RDKit ETKDG-only and ETKDG + MMFF workflows exhibit steeper empirical scaling trends of approximately T proportional to B^3.04 and T proportional to B^2.99, respectively. These results indicate that the proposed bond-centric reconstruction strategy becomes increasingly advantageous for large polymer-scale molecular graphs. Overall, these results demonstrate that bond-centric local geometry prediction provides a scalable and physically consistent approach for rapid three-dimensional molecular structure generation, offering a practical alternative for high-throughput molecular modeling.