What you say has nothing to do with TFA, which is not about ML-KEM but about the session key establishment protocol used in TLS, in which ML-KEM is just a component.

DJB supports the use of ML-KEM in TLS, but he correctly says that using only ML-KEM is unwise, because absolutely nobody can guarantee that no method to break ML-KEM will be discovered in the next years, as it already happened with the algorithm that was preferred before ML-KEM, until it was broken a few years ago.