While I don't know how exactly the FA2 works or what is the exact reason for the worse ADEV value when external reference is used (I can also confirm that, checked with T-adapter), but my guess for the root cause is that the external reference is connected to a different pin of the Max II CPLD. Either the routing to that pin is noisier or the routing within the CPLD is noisier.
When measuring own reference most of the phase noise effect of the reference is taken out from the measurement. This is because when the interpolating part is measured (the time difference between the signal edge and the next reference edge) if the signal = reference (maybe with some shift) even if there is phase error, the phases for signal (= reference) and reference are moving together.
Thus when measuring own reference most of the phase noise of the reference is eliminated. That is what we're measuring is the noise of the measurement setup itself.
So while I maybe wrong with this whole theory, this means that the measurement setup for the external reference is worse.
This is unlucky, since in case of the external reference input a better precision reference can be used. However in the light of this result it seems that with external reference only the accuracy can be improved on the price of loosing some stability.
The good thing is that FA2 is still pretty good even when using external reference.
One way to cross check this theory is temporally removing the OCXO from its place (or maybe its signal output only to provide for the rest of the circuits the same load) and directly connecting an external reference there.
Some comments to older posts.
The 190823 board has a 6 pin IC and the other board did not have it.
True, but even the IC is in place, some DC blocking capacitors are not (missing also from my PCB) so net result is the same: That block probably planned to do some input signal shaping is not doing anything even in case of my variant of the board.
The delays probably calibrated and stored in the EEPROM. The question is: are the devices individually calibrated during production or was a single characteristic recorded once for all?
Probably not. In the FPGA solutions typically much better precision is aimed with a single measurement. Thus in FPGA case it matters a lot that the thermometer code is linear. In case of FA2, in statistical mode probably multiple measurements are done during the measurement period and the expectation is that even the thermometer code is not linear it averages out.