JPEG XL decoder no longer performs 'quantization constraint', it is a classic loop filter, just with heuristics and control fields to maintain detail much better than loop filters in video codecs.
If you start with sharp high quality originals that don't have yuv420 in them, you usually will see 65 % savings with JPEG XL, i.e. more than 30 % on top of brunsli. You can think of VarDCT-mode as guetzli(-35 %) + brunsli(-22 %) + format specific changes(-30 %) like variable sized dcts, better colorspace, adaptive quantization and loop filtering.