<p>Parametric computer-aided design (CAD) plays a central role in modern engineering workflows by enabling explicit design intent representation, model reusability, and efficient downstream modification. However, automating parametric CAD modeling from natural language descriptions remains challenging, particularly in maintaining structural consistency and generating compact, executable modeling scripts. Existing code-centric autoregressive methods may suffer from fragmented geometric planning, while diffusion-based approaches face efficiency limitations and difficulties in handling discrete CAD representations. This paper proposes Lite-AR-CADecoder (LACAD), a lightweight framework for text-guided parametric CAD modeling in the CadQuery environment. LACAD targets executable script-based parametric CAD generation in the CadQuery environment, with emphasis on executability, structural correctness, and parametric fidelity under the proposed evaluation protocol. The framework incorporates a structure-aware autoencoder that jointly encodes abstract syntax tree (AST) information and sequential command features to construct compact structure-aware latent representations of parametric CAD programs. A flow-based latent generation module models the mapping from textual descriptions to the learned latent program space, while an autoregressive decoder reconstructs executable CadQuery instructions from the generated latent representations. To support systematic evaluation, a new CadQuery-based dataset with verified executable scripts is constructed. Experimental results show that LACAD achieves competitive command-level and geometry-level performance compared with diffusion-based and autoregressive baselines, while using only 230&#xa0;M parameters. These results suggest that lightweight structure-aware latent generation is a promising research direction for executable text-guided parametric CAD modeling.</p>