Do tx kernels have to be kept forever?

from my initial (mis)understanding of the MW paper, I thought the main advantage is that transactions can be merged together - a whole block becomes essentially one transaction, resulting in massive scalability.

Now I see that for each original transaction, we still have to keep its kernel. So the scaling as compared to bitcoin is only a constant factor (something like 2.5x) . Is this correct?

Or can transaction kernels eventually be pruned somehow?

yes, current secp256k1 kerners must be kept forever.

the small factor savings over bitcoin becomes an order of magnitude if bitcoin were to adopt confidential transactions.