Finally, distributors of consumer wallets can use this research in
order to distribute their wallet with policies which may be less prone
to Tor-specific attacks. Or leave this out altogether if their
audience has different expectations for connecting to Bitcoin.

Sure. I guess there will be wallets for all kinds of people in future, sharing a common core that they can customise (this is certainly the vision and general direction for bitcoinj, and it's working out OK).

To clarify, my comments above were for mainstream granny-focused wallets. Wallets designed for crypto geeks can and should expose all the knobs to let people run wild.
 
One possible direction to go is to use Tor for writing to the network and use general link encryption and better Bloom filtering for reading it. Thus new transactions would pop out of Tor exits, but there isn't much they can do that's malicious there except mutate them or block them entirely. If you insert the same transaction into the P2P network via say 10 randomly chosen exits, the worst a malicious mutator can do is race the real transaction and that's no different to a malicious P2P node. Even in a world where an attacker has DoS-banned a lot of nodes and now controls your TX submission path entirely, it's hard to see how it helps them.

The nice thing about the above approach is that it solves the latency problems. Startup speed is really an issue for reading from the network: just syncing the block chain is already enough of a speed hit without adding consensus sync as well. But if you're syncing the block chain via the clearnet you can connect to Tor in parallel so that by the time the user has scanned a QR code, verified the details on the screen and then pressed the Pay button, you have a warm connection and can upload the TX through that. It reduces the level of startup time optimisation needed, although Tor consensus download is still too slow even to race a QR code scan at the moment. I think tuning the consensus caching process and switching to a fresh one on the fly might be the way to go.

When BIP70 is in use, you wouldn't write the tx to the network yourself but you could download the PaymentRequest and upload the Payment message via an SSLd Tor connection to the merchant. Then malicious exits can only DoS you but not do anything else so there's no need for multiple exit paths simultaneously.