We unconditionally use the portable implementation of SHA3 in some places in ML-KEM, e.g.: - one-shot hashing on AVX2 - secret key serialization We should rule out that this has adverse effects on performance, stack usage or binary size.