Ok, as per your PM I'll just say the following. Bjornl is correct is most of what he is saying, and might in fact be 100% correct based on his experience. His experience or experiences do not necessarily correspond to everybody else's AND what what works with one platform or cpu, or even the exact same configuration, is not always mirrored on the next persons attempt.
I just recently went through this with both a build I did for a client AND a member who posted a help thread here at Tom's. While the information below only actually applies to the Tom's thread, the reason for the trouble they ran into and the end result were approximately the same. This has always been a factor, on ALL platforms, since forever. The fact that the modules are higher clocked is not a problem though. They will either run at their supposed configuration speeds or they won't. So long as they DO work together, you can try to tune them in using minor adjustments if they don't want to run at that speed and worse case scenario is that you end up running them a slightly lower speed but this is not bad, as memory designed to run at higher speeds will almost always run at any lower speed with less voltage and therefore will run cooler and last longer. I always buy memory that is capable of running at a higher speed than what I intend to run it at, when buying it for myself.
It's true that a lot of the time you CAN mix and match RAM and so long as you keep the module selection limited to models with fairly similar or nearly exact specs, they will play nice together. Even, sometimes, completely disparate modules will do so with one module having it's primary settings adjusted by the system to match the other, lower clocked module. This is NOT always the case however AND instances of memory that are not identical or in fact even MATCHED and TESTED to work together, not working together, have been on the rise since the introduction of DDR4. At least from what I have seen through my own experience using them on builds and from what I've seen here in addition to what I've read at a number of different places including reviews and other forums.
The biggest problem, I think, is that companies like G.Skill, Corsair, Kingston, etc., do not actually make the memory chips used on their modules and they tend to source those ICs from a variety of different places, even on some of their highest end offerings. Buying another memory module later on that is the EXACT SAME part number is NO guarantee that you will be getting either an identical module or even one that is close to the same. For example, in the image below you can see that although the two lower pictured items have the same exact part number, having likely been part of different production runs that completely changed up the configuration of the modules, they do not use the same size IC, the same brand of IC, the IC are ranked differently and one module is single sided while the other has ICs on both sides.
None of these things SPECIFICALLY means that these modules won't work together, howerver, these are THE EXACT SAME part number, so you can imagine what kinds of differences there might be between two modules that are entirely different part numbers even though they might be the same series and manufacturer. What it does mean, while there is no absolute condition saying these can't play nice together, is that for every factor that is different there is an increased possibility that for whatever reason (wildly different sub-settings, voltage requirements, or just plain oneriness) they may require advanced tuning or in the end may simply refuse to work together regardless of configuration settings. It happens. Nothing Bjornl or anybody else says can change the fact that sometimes modules simply WILL NOT play nice together and the further you stray from "completely identical" the higher the probability is in most cases that this will be so. Again, that does not mean that in some cases you won't be able to use two wildly different modules together. That happens too.
But there are significant headaches involved in trying to get two modules to run, either together OR in dual channel operation, when they don't want to. Sometimes the resulting headache includes having to go through the hassle of tuning only to find out that it ain't happening and you now have to go through the additional hassle of trying to return the memory which involves generating a return, getting it approved, waiting while it's shipped and received so you can get a credit and then try another part number. And after all that, when you get a different module, you are still at the mercy of a potential crapshoot because the next module or part number might not have any more success than the first one did. Of course, you might get lucky on the very first attempt as well or there may simply be no issue at all and the modules or motherboard you have may be very forgiving of differences between modules.
We have very clear and compelling evidence to show that the CPU and motherboard sometimes play a role in this as well, but that's a whole other conversation. Furthermore, and this is just an impulse opinion which is not fact but is based on some experiences in the past with enterprise level hardware, it seems entirely likely to me that Threadripper and X399, which are at least semi-enterprise hardwares (Using hardware from the Epyc skus. Some will of course argue that these are not actually enterprise hardwares, but that too is a separate argument.), not enthusiast or mainstream hardwares (At least, not strictly, by intent), are probably somewhat more likely to adhere to a somewhat different standard that the mainstream/enthusiast Ryzen (3, 5 and 7) lineup since it also supports ECC and quad channel operation (Using TWO dual channel memory controllers) as well as having a Non Uniform Memory Access architecture which Ryzen 3, 5 and 7 do not share.
NUMA is a completely different can of worms involving too many factors to explain or argue here, but suffice to say that it the platform AS A WHOLE is significantly different than what is used by it's smaller brethren, enough so that I believe direct comparisons are unfair and probably not entirely realistic. What works on Ryzen 7 may not actually be the same as what the best practices are when using Threadripper or Epyc. I'm sure they probably share a lot of similarities too and I'm not claiming to be an expert on ANY of these hardwares, only a systems builder and enthusiast who knows enough to steer clear of as many potential complications as possible, whenever possible.
To that effect, I feel the same as most others here and as our previously resident memory expert Tradesman1 did, which is that if you are going to buy memory you should make every effort to buy the amount of memory you every plan to use from the start, and do it by buying them in a matched/tested set that uses the fewest possible number of modules needed to obtain the desired total capacity. This not only greatly increases the probability that your memory will be a trouble free component of your build but it also allows that later on down the road, if in fact you discover at some point that your memory capacity is NOT enough, you still have room to add more.
Yes, you WILL be facing THIS exact problem we are discussing if that eventuality comes about, but at least you will have the option of doing so without having to buy ALL of the memory over again in order to obtain higher capacity sticks. You can take you chances, or you can be fairly certain that what you are purchasing will work without any added complications. It is worth noting though that in some few cases, there have been instances of even memory purchased together in a matched set not playing nice together despite supposedly having been tested as compatible, so even if you do so there still remains an element of uncertainty ANY time two modules are used together but it's a necessary evil since dual channel operation cannot be achieved with a single module.
Again, it's worth noting that I am NOT disputing anything Bjornl has said, only expanding on it and offering a wider view of the picture in general.