They assemble one o top of the other.
The board have pin 1 and pin 2 swapped on both LTZ1000 ... Kicad played some bad trick when it lost the symbols reconnecting it in reverse order when I reassigned it. Anyway I managed to not notice the problem despite the many times I have checked the schematics. I had to bodge some fix that fortunately seems to have no much impact on thermal design.
The board had also some filtering if you wanted to use autozero opamp for the LTZ reference. This seems to cause more problems so I just reverted to original LTZ circuit bypassing filtering.
Once fixed board I had to build a TEK mini thermal chamber to properly test TC ... and this has taken me some more time.
Averaged reference seems pretty stable and also TC looks ok ... you'll never know how it works until you do not put inside the 3458a.
Then the project was suspended because I was busy moving to a new house ... things are moving slow now.
The upper board hosts the two LTZs.
The bottom board hosts connection to 3458a, averaging/buffering circuit, and virtual ground generator for second zener.
There are some unconnected copper pour (both sides) just aiming to improve temperature equalization around the board, don't know how it improves ...
Virtual ground is used for second zener, referencing pin 7 of first LTZ.
I used only one divider shared between the two LTZ for temperature setting, "powered" from the averaged/buffered reference. Sensing of the averaged buffered out reference is done at the 3458a connector pin, so what it goes to the divider can suffer from some copper TC. I made some raw calculation and the effect should be very minimal.
PS: Routing the board trying to keep EMF balanced was a hell of work ... I can see why you selected ceramic substrate.