I guess I meant losslessly round-trippable. In other words, you can go from jpeg -> jxl -> jpeg without any loss in quality, potentially (although with jxl -> jpeg -> jxl, you will lose space while it is a jpeg, and you'd probably have to pick a high compression quality in order to not lose information... you may also lose information such as metadata that jxl accommodates but jpeg does not, like transparency)

So backwards-compatible in the sense that the jpeg-xl algorithm spec can read jpg and store the same pixel data more efficiently as jxl if you like. You gain space and lose nothing (except perhaps encode/decode speed).