Generating molecules that simultaneously achieve optimal 3D pocket-binding conformations and chemically plausible topologies remains a central challenge in AI for Structure-Based Drug Design. Graph-based models excel in SE(3)-equivariant spatial reasoning but often struggle to ensure chemical validity, whereas language models capture discrete chemical syntax yet lack 3D spatial understanding. Here we introduce SE3-BiLingoMol, an SE(3)-equivariant language model for pocket-aware 3D ligand de novo generation and fragment-guided optimization. Built upon Geometric Algebra Transformers and a fragment-aware SMILES representation, our model enables SE(3)-equivariant modeling of continuous 3D geometry while ensuring chemically valid molecular topologies. To counteract cumulative 3D conformational errors inherent to autoregressive generation, we developed a bidirectional attention-based self-refinement mechanism as a key architectural component of SE3-BiLingoMol. Our model achieves state-of-the-art performance in an in-silico evaluation across over 100 diverse targets. Critically, application of SE3-BiLingoMol led to the discovery of a novel tetracyclic HPK1 inhibitor showing potent in vitro activity and robust in vivo anti-tumor efficacy. This work demonstrates a powerful and practical generative AI framework for accelerating structure-based drug design.