Some thoughts on that:
Why not use torrent code outright? That way, lots of code reuse could happen. A node could be a full or a partial node (perhaps a configurable %age of blockchain?). A full node could act as torrent seeder. Since the blockchain evolves, how about simulating a file for each x blocks/MB that is then loaded via torrent - would also speed up initializing a node (instead of requesting blocks one-by-one). File compression could also be considered here.

1% of blockchain with 100 nodes would probably be problematic, each piece of the blockchain needs to be held by multiple nodes in case a node drops out of the network. Even when only two nodes have unique information, they might happen to belong to the same network and get lost when going offline.

How would a partial node know which blocks it needs for having complete info on the addresses it holds without broadcasting its addresses (thinking BitMessage here)? Would it download each block but only keep those it needs? Then it would still take ages for a blockchain like Bitcoin to first initialize. Use something like Tor to anonymously find out about the blocks required, but then download those (and some random blocks for obfuscation) via "normal" network?

Store the info about blocks needed in wallet.dat as a way to "cache" it in case of a required blockchain-re-download?

How to handle the improbably event of a new address generation yielding an already existant address? re-scan for needed blocks after any address generation? Also would be necessary when importing an existing address.

Kind regards
Mike