High-fidelity simulation of large, slender structural components, typical of composite and other heterogeneous materials, is often limited by the cost and robustness of linear solvers. Accurately resolving thin structures requires highly stretched finite element meshes, which lead to severely ill-conditioned linear systems. In this setting, iterative solvers may stagnate, and even Krylov methods coupled with advanced algebraic multigrid preconditioners do not consistently outperform direct factorisation . Over the last decade, scalability has been addressed using multigrid and domain decomposition methods with enriched coarse spaces, such as GenEO. These methods aim to capture the dominant global behaviour of the discretised problem at a coarse level, allowing local smoothing and subdomain corrections to efficiently reduce the remaining error. When the coarse space properly represents low-frequency error and inter-subdomain coupling, near mesh-independent convergence can be achieved. However, for extreme discretisations combining very large element aspect ratios with strong coefficient contrasts, standard coarse spaces often fail to capture essential global modes, leading to degraded robustness and slower convergence. To address this limitation, we propose a two-level preconditioner tailored to extreme discretisations. The coarse space explicitly separates stretch-dominated modes aligned with the mesh distortion from transverse components, enabling a more accurate representation of the global response. As a result, convergence depends only weakly on the mesh aspect ratio and overall domain dimensions. Elasticity benchmarks in composite materials confirm improved robustness and efficiency on strongly stretched meshes.