Over complicating the build system is a reason for dropping almost anything but especially this. The whole point of a framework like that is to make things easier.

That's the one criticism I found valid. I even called it one of the "Four Horsemen of Poor Performance" in a pretty widely read article I wrote on server performance back in 2002. On the other hand, over-reliance on static allocation means allocating many arrays/pools for the worst case, even when they couldn't possibly all hit worst case at the same time, and that can be a pretty bad choice on a highly memory-constrained device. (Doing it for the sake of real-time predictability is actually a different issue.)

OTOH complaining about not using templates because of performance seems exactly backwards, and complaining that it's not likely the the project sponsor will port already-open-source code to a competing architecture themselves seems a bit too entitled. When I see a list of negatives that are mostly bogus, and no mention of positives at all, it's a strong indicator of NIH syndrome.

It is confusing though as the code mostly appears to be Apache, not sure if it is just parts that are constrained by the EULA, but it is clearly not intended to be ported to RISC-V or MIPS or whatever.

You don't need any EULA to use any of Mbed's stuff, it's all Apache licensed and can use any compiler.

I have not used it but zephyr (https://www.zephyrproject.org) is a cross platform (including riscv) os/platform that looks interesting for embedded development.