Reproduction of Hy-Embodied-0.5-VLA (arXiv 2606.14409): flow-matching vision-language-action model; action-chunk reconstruction PoC.