4MPts isn't exactly deep indeed but the way (-I think-) Keysight works it is easely expandable to 40Mpts, 400MPts, 4Gpts. Doesn't really matter.
Well, no, for a couple of reasons. We'll see about that.
And Keysight doesn't sacrifice anything. If I'm right they use two parallel acquisition systems.
For sure they sacrifice functionality for update rate. E.g. the first thing they sacrificed was memory depth. In the past they also crippled measurements by using only the onscreen data. Someone in this forum stated this is no longer the case, since they introduced a measurement buffer but my impression is that this is also some kind of workaround instead of measuring and displaying from the actual sample buffer as LeCroy and Tek scopes do.
But anyway: I think it's clear that they only achieve the high update rate by using an expensive specialized ASIC and this makes it impossible for them to easily increase the memory depth. So whatever their current approach is, it can't be easily extendend to 4GPts.
One with the large memory which works like a continuous circular buffer and one with a short buffer (enough to fill the screen) to do high update rates. Once you press stop Keysight scopes wait until the rest of the large memory buffer is filled before allowing the user to scroll left / right. It is a very simple & elegant solution.
Actually that doesn't seem to be possible even from a theoretical point of view. If you're in normal trigger mode and press stop, there is usually no way to continue sampling from the trigger point unless pressing the button happens more or less at the same time as the last trigger event. E.g. assuming the current DOSX3000T Keysights have 4MPts for your hypothetical circular buffer: at 2.5GSa/s, it would be filled within 1.6ms. So if you stop the aquisition in normal trigger mode and the trigger event happened e.g. 10ms before, the circular buffer was filled again multiple times. Sorry, but this just doesn't seem to be a feasible approach.
Again, what you're suggesting is IMHO only possible in single trigger mode and looking into the DSOX3000T manual seems to confirm this. Quote from the manual: "Pressing [Single] instead of [Stop] fills the maximum memory depth". And I think that other scopes fill the whole selected ("maximum") memory in single trigger mode as well (if this is feasible).
Besides, even letting all of this aside, your suggested approach would not be "easily expandable" to any memory depth. As you say, the scope needs to fill the buffer before it lets you have a look, scroll around etc. Now with 4Gpts and a sample rate of 1MSa/s, it would take 4000 seconds to fill the buffer. That's more than an hour. Now with the 4Mpts of a current Keysight, that's not as much of an issue unless you select a 1kSa/s or so (if that's even possible). Actually I would think that a Keysight doesn't allow you to select an acquisition setup that would delay reaction to a key by even a few seconds. Yet with 200MPts of a Siglent SDS2000X Plus, it would take 200 seconds at 1MSa/s to fill the memory. That doesn't sound horribly feasible either. So it's natural that a scope tries to select either memory depth or sample rate to avoid getting locked or slow down to a crawl. Actually, of all scopes, LeCroys tend to allow you setting up acquisition modes that are terribly slow or only fill a part of the screen and Keysight adepts criticize them for exactly this lack of automatically selecting a setup that gives immediate feedback.