A JPEG image is divided into tiles (called MCUs - minimum compression units) and each tile is processed separately. The size of the tiles is 8x8 pixels, but JPEG can save chrominance (hue) information at smaller resolution and therefore a 8x8 pixels of chrominance information can effectively affect for example 16x16 pixels.
When saving again, the tiles are processed again, which can lead to further losses. There are software systems that copy unmodified MCUs 1:1 into the modified file without optimizing them again. However, this does not work with all JPEG formats.
Note: Lossless JPG crop works better, if the selection coordinates/dimensions are a multiple of 16 (JPG specification). If this is not the case, the result image may have slightly different dimensions or the start point.
So it would seem that it does cropping of existing compressed image without decompressing but you may have different dimensions or start point if you do not choose dimensions that are a multiple of 16.
This indicates to me that lossless crop will be restricted to dimensions and start point that are on the boundaries of the compression blocks.