A deeply considered take: this is theoretically interesting but in practice matters very little.
It’s like worrying about compiler bugs instead of semantic programming errors. If you are getting to the point where you have machine checked proofs, you are already doing great. It’s way more likely that you have a specification bug or modeling error than an error in the solver.
If you are worried about it, many people have solutions that are very compelling. There are LCF style theorem provers like HOL Light with tiny kernels. But for any real world systems, none of this matters, use any automated solvers you like, it will be fine.
It’s like worrying about compiler bugs instead of semantic programming errors. If you are getting to the point where you have machine checked proofs, you are already doing great. It’s way more likely that you have a specification bug or modeling error than an error in the solver.
If you are worried about it, many people have solutions that are very compelling. There are LCF style theorem provers like HOL Light with tiny kernels. But for any real world systems, none of this matters, use any automated solvers you like, it will be fine.