zarr mostly just makes the tiled layout explicit, but processing tools still need to be tile-based.

I think I have to disagree here. It allows in particular for storing additional metadata which can be ingested by xarray, for example coordinate transformations. Furthermore, zarr combined with OME-NGFF also provides a very standardized way of storing additional data other than the image data. For example, currently label data. The SpatialData framework is ahead at the moment of the OME-NGFF spec and already supports points, shapes and tables as well which is looking to be included in OME-NGFF too.