One example gmaxwell gives: all presigned nlocktime transactions would be broken. For users keeping these in storage they may well represent a lot of security. Gone... the moment a new version of the software no longer sees the transaction as being valid.
As far as I see it, if malleability can be fixed in such a way that older versions of the software still see immalleable transactions as valid transactions then, well
do it.
In a soft fork, by definition, the new version of the software can reject transactions that the revious version considered OK.
For example, IIUC the soft-forked SegWit proposal implies redefining an op code that previously meant "no-op" to mean "check the signatures in the extension record" or something like that. Thus, a transaction that used that opcode (for some bizarre reason of its own, possibly fraudulent) could be valid before SegWit was enabled, but become invalid after it.
That may be a good argument to phase out nLocktime in favor of CLTV. Once a transaction is in the blockchain, its position in it defines the rules by which it should be validated, which allows proper handling of old time locks.