I always thought it was because of the added cost from increased design complexity. Is it something else?
The complexity is real, but in a typical SoC the actual CPU core is maybe 10% of the area and tossing in an MMU impacts maybe 1-2% of the total chip area. I haven't seen the pricing sheets, but I suspect the much bigger cost is the higher royalty payment associated with instantiating the MMU.
There's nothing stopping you from taking say, a Raspberry Pi 4 and using it as a bare metal device (no OS) just like an Arduino. Or taking that same chip (BCM2711) and putting it in your own board to do something similar. It just wouldn't be economical for that sort of purpose most of the time.