I assume that Q4 and Q6 are substrate PNPs and the yellow-brown stuff around their emitters is just bulk epitaxial layer silicon with no buried layer underneath (buried layer seems to show up as red-brown).
I learned that this is wrong and that buried layer is not visible externally. The only sign of its presence is a slight vertical depression on the surface of the epitaxial layer and even that is shifted vertically with respect to the real location of the buried structures. This phenomenon is known as "buried layer pattern shift" and a lot has been written about it because it's a nuisance for IC fabs - it makes it difficult to align surface features to the buried ones.
maybe I'm wrong and it is actually a P diffusion which acts as the GND collector and it covers almost the whole island and makes contact with P isolation diffusions surrounding it to rid itself of the incoming current
This appears to be the case. The mustard colored areas at the perimeter of isolation islands have to be P diffusions produced in the same step as NPN bases. I really tried to find some confirmation in literature available online and failed, except for this somewhat ambiguous suggestion in "Analysis and Design of Analog Integrated Circuits":
The terraced effect on the surface of the device results from the fact that additional oxide is grown during each diffusion cycle, so that the oxide is thickest over the epitaxial region, where no oxide has been removed, is less thick over the base and isolation regions, which are both opened at the base mask step, and is thinnest over the emitter diffusion.
Additional evidence is the fact that the long green N+ silicon resistor on the LM723J die appears to overlap slightly some isolation islands and yet makes no electrical contact with them, so it needs to have some grounded P silicon below it for isolation. Also, the MIL723 die has places where a P resistor connects an NPN base with the perimeter of the isolation island and it all looks perfectly smooth.
Armed with this knowledge we can have another look at the National LM723J and its D2 diode.
On the right, we see the emitter and base of Q6, covered by metal and surrounded by "mustard", which has to be a surface diffused collector extending all the way to the isolation diffusion around the island. Above Q6, an N+ diffusion is made into the collector and part of it branches off to the right and crosses two vertical metal traces. I thought that it connects with the metal on the far right, which is GND, but with some magnification of the image it becomes apparent that it actually does not.
Since the N+ diffusion is fed constant current from Q3, it forms a zener diode with the grounded P silicon below (D2 on the National schematic) and assumes a fixed potential of a few volts above ground. So what's the point of the stub extending to the right? The stub is just long enough to leave the isolation island of Q6 and reach an isolation diffusion, which has higher concentration of P dopant than Q6 collector. Higher doping is known to reduce breakdown voltage (see Linear AN-82, buried zeners chapter) and I presume it's the reason why D2 has lower breakdown voltage than D1, which is a plain BE junction.
The metal connection to D2 cathode branches to the left and connects with structures in another isolation island. It appears to be connected to the bulk of the island and a P diffusion in it. The P diffusion has a comb-like N+ diffusion inside, forming a base-emitter diode. Reverse bias across this diode is no more than 5.9V (VREF minus two diode drops) and almost zero if we assume that D2 voltage is 5.7V. Therefore the chance that this diode forms some super-fancy large area zener appears rather slim. Furthermore, such zener would be in series with the wimpy D2 zener, so what's the point?
I think that the "comb" really is a compensation capacitor and nothing more. This is a bandgap chip, plain and simple. The datasheet schematic checks out 100% and it all makes sense.
Let's do the loop gain. Rising Q21 Vbe by 3mV we get 3µA extra current in Q21 and 1µA in Q22. That's 4µA total and 48mV extra across R15. So for a 51mV rise of Q21,Q22 base wrt ground we get 3µA rise in Q21 and 1µA in Q22 and therefore -2µA at the Q21,Q19 junction and a -2V shift of Q6 base voltage and the VREF node.
That's not a lot
Yep, loop gain is only some 20~40dB. But current through Q5 is almost 0.5mA due to the resistive divider, so Q5 emitter output impedance is some 50~70Ω depending on temperature. Then, closed loop output impedance is a few ohms or maybe a bit under one ohm. I can believe it.