The consensus already assumes 2/3 honest validators. Why do we need to use redundancy and go through so much trouble in block generation? Isn't DA guarantee already inherent to the consensus protocol?

We want to ensure that minimum trust is needed on the DA layer. For the consensus, we assume the super-majority of validators to remain honest and make sure the blockchain progresses and finalizes. However, we want the applications using the DA layer to not have to participate as a validator or full node to ensure that the data is available. Only a very limited amount of data querying should provide enough confidence for the applications to keep using our proposed layer. This is why we modify our block structure and make sure that a block header with a constant number of data queries is sufficient to provide a ~100% guarantee that the entire block is available. We emphasize the fact that none of these guarantees about DA assume any dependency on any other honest participant, as far as availability is concerned.

Also, the applications can choose to keep only parts of the block which are relevant to their app. If enough such applications are present in the system, it ensures that collectively the entire block data is available.