I don't really see how this is a problem: if a part of the "joined" transaction becomes invalid, I assume the coinjoin can just recreate a new transaction without that input and broadcast it again, right?
Maybe. Or maybe a malicious attacker (maybe some other coinjoin wallet who wants to attack their competitors) is flooding your coinjoin service with inputs that they are double spending, so it happens again on your next attempt, and your next one, and your next one, rendering it impossible for you to coinjoin any coins at all. The same could apply to any custodial Lightning wallet, as another example.
While I do acknowledge the problem here, I don't like that a possible attack on CoinJoin or Lightning transactions can be resolved only by enforcing a rule on all the transactions.
But it is neither a rule nor being enforced. This is a change with the Bitcoin Core software, not the Bitcoin protocol, giving node operators the easy option to accept replacements for all transactions in their mempool. Node operators can already do this if they wish, and some alternative full node software (such as Knots) has had this option for years already.