Yeah, I very much expect the manual element to the "baking" process would have been a human sanity check on what algorithmic settlement placements the procgen spat out, and "no more", and even then you can only spend so many manhours doing it -- there's plenty of peculiar locations that must have been "ok"-clicked past on "idling engine", by some poor, tired employee on QA duty.
The one I keep bringing up is Mic Turner base over in the California Nebula, which was originally built to study a field of "space pumpkins" there, and as such located right next to it, back when "flora" was added to the game. A tourist beacon was subsequently added, recounting this, but today that beacon, and the field it's in, is on the opposite side of the body, to the base, and in a ravine, rather than a flat area, and somehow a Thargoid barnacle has ended up inside one of the large landing pad slabs at the base.
Now, I'd like to think that the latter is a deliberate bit of environmental story telling, to create rumors about what Mahon and company may be up to, but the likelyhood of it
simply being a procgen edge case, overlooked by QA, rings ever so much truer to jaded ears. Already in Horizons, the type of flora in the field had changed, at some point.
I am also pretty sure every settlement and other POI comes with a bitmap that modifies the ground to fit the prefabbed environment - maybe with just a single multiplication, I don't know. It looks kind of funny in the SRV intro tutorial, where it produces an isolated plateu down the slope of a crater wall, demonstrating that it can brute-force some rather extreme cases, and doesn't necessarily need hand placement to make something "usable".
Probably, something something going wrong with the application of this, between visible and collision geometry, is the cause of the occasional place where you find your SRV driving on an invisible surface above ground -- there's also the occasional handful of surface scatter rocks, whose rendering does not become suppressed, in a landing pad shaft.
Balancing noise in the new terrain generation would reasonably be about tweaking factors and limits in procgen rules, regardless of why one may feel the need for balancing -- I'm kind of wondering whether we're talking an algorithmic noise function actually generating new pseudo-random noise every time, or if it is just as much one or a few pre-generated bitmaps of static, as all other terrain "tiles"; Unlike them, with noise I'm pretty sure I'd rarely notice - no interesting features that stand out and catch your eye.
Oh, and extrapolation is dead reconning the tail of a trend.