I did some research on hash collisions today and there are collisions in MD5 that have 2 inputs but the same output. In sha256 or even sha128 you will never have the same output. I am also aware that a seed of 12 words will surely be unbreakable for the next 50 to 100 years. I think I heard that it is possible to access coins in the wallet with different seeds, but probably with subaccounts.
To be clear hash collision (even with weak algorithms) has nothing to do with finding a mnemonic collision.
The hash algorithms that are used under the hood of BIP39/BIP32 are there to give us a deterministic way of deriving child keys from an entropy and the security of this whole setup is determined by the size of that entropy not by the collision strength of the underlying hash algorithm. In other words in a 12-word seed you still have to face 2
128 possibilities even if the scheme was using a hash algorithm that is weak against collision attack like SHA1/MD5 (ignoring their small digest size).