No, transaction inputs are specified in terms of txhash:output_index pairs, not addresses. You don't import value from an address, you literally import it from a particular transaction.
The pruning process isn't implemented today, but when it is there are a few possibilities. One is that the tx won't be completely deleted, the hash of it will still stick around in the database. It's 32 bytes instead of several hundred, so it's still a big win. Another possibility is that some nodes will keep the whole chain and if a tx input isn't known, pruning nodes can contact archival nodes and ask them for the tx. If they don't have it, it's an orphan transaction. If they do, it means the connected tx was pruned and thus it's a double spent.
You're saying that the option that the tx won't be completely deleted is just one of the possibilities. What I am very interested in, is whether there is a possibility that previous transactions can be completely deleted, as is claimed in Nakamoto's paper.
After studying the protocol, for now my conclusion is that this is not the case. Referencing a former transaction instead of an account unfortunately doesn't make a difference. Could anyone please invalidate my conclusion?
Honestly, I'm skeptical pruning will ever really be needed. Assuming continued growth, disk storage is so cheap that the cost of keeping the old parts of the chain around will be dwarfed by the costs of keeping up with the head.
Whether pruning is necessary or not is another question, but I'm glad you brought it up. Assuming only the transaction hash sticks around like you say and 2000 transactions per second as mentioned on the scalability page, this means a storage need of 32 * 2000 * 3600 =5529600000 bytes = 5.5 Gb per day. Not to mention that this must be stored in such a way that it is searchable quickly, which will likely increase storage needs.
All and all, quite a lot more than the 'best cast' of 80 bytes per block (or 1.16 MB / day) that is mentioned on the scalability page. I hope you understand why I'm so damned curious about this.