I have just been working with some Esri vector tiles, which support a variable zoom depth. That is, tiles are generated to higher zoom levels in areas that contain more detail, and aren't generated in comparatively empty areas.

https://developers.arcgis.com/rest/services-reference/vector-tilemap.htm

It has obvious benefits when statically generating large tilesets that cover both urban and rural areas. The downside is that you also need to communicate to the client where tiles are available. Esri does this through a hefty tilemap JSON file, but there probably more efficient ways.

Is it worth putting something like this in TileJSON? Or maybe it doesn't make sense if Mapbox isn't going to support it.