Good luck certifying an automotive or an industrial system that can lead to loss of life or severe disruption if it malfunctions. As far as I am aware, QNX, VxWorks, and CMSIS have all been used in systems that passed certifications. As IoT &co. get applied more broadly in our society, they are put in more places where they are expected to function flawlessly. I would rather want my power grid or water utility operator to use CMSIS, for example, than rt-linux.
The only serious user of rt-linux which uses it to build a hard-RT system that I know is SpaceX. Look up the Reddit AMA post from them on how rt-linux is used. They are staying on a specific kernel version, had to do kernel patches (other than applying rt-linux), they use a very specific model of an old Xeon for low latency interrupts, incorrect programming of one process can disrupt the RT behaviour of another process, and most importantly, they have a secondary board, a truly real time one (PowerPC based, IIRC) to communicate with the real hardware.
One interesting morsel in all of what you described is that for all of the hoop-jumping SpaceX does to bolster rt-Linux, it is -still- easier/faster/ better by some business metric for them than the alternative.
If the redundant hardware and some ground rules they have in place have made the onboarding of developers better and faster, than perhaps it should be recognized that they seem to be prioritizing the DX - especially since plenty of applications in RT will also use redundant hardware and confine their hardware selection similarly to prioritize the performance targets of their application, SpaceX just added developers to the feature matrix.