I'm curious what you've heard about them that suggests this?

As far as I can tell the maintenance overhead to mapping all the available type from/to python as nicely usable python APIs is a major pain and not just because of cythons current implementation but of e.g. conceptually mismatch between "good" python and "good" C++/Rust/whatever implementations. I.e. you could say that you have a paradigm mismatch between the underlying implementation (of whatever you extend) ad the python API which makes the already hard problem of keeping an API stable even harder.

Furthermore tooling to work in such contexts (as a API consumer) often is very limited or does not exists, but the API producer seldomly has the time/money to produce such tooling either.

Lastly there is the remaining problem of Python by itself being really slow, for it's main use-cases this often doesn't matter as e.g. the scientific computations are run by native extensions python delegates the work, too. But this would mean that you have to also support native extensions support or hope that no extension needs to be really fast (not needing to be really fast is surprisingly often the case) or similar solutions. In turn choosing something which has a good JIT, AOT or similar compilation has benefits.

I think the worst thing about Python and Blender was not necessarily the choice of language, but the frequent API changes and not having anything close to an IDE-like environment.

Not only did old examples often just not work, there was no serious code completion to facilitate creating anything from scratch.

Exactly this. One of the reasons Maya remains favored in studio environments is its strong API stability guarantees. Python is excellent for automation when performance is less of a concern.