I thought the whole point of the Generate Assets feature was the multi-resolution output, but apparently the only input data the exporter makes use of is the layer’s raster data, and not the original resolution of a linked or embedded smart object, thus making any upscaling from a 1x version unfeasible.
This way the ‘retina’ versions of the image are simply created as blurry memory-intensive pixel-bloats of the original size, which is nothing anyone could want 😞