I had the same gut reaction. It may at least be possible, if extremely difficult, to get past the need for good repute ... LN penalty not ideal here, clearly. 

 I don't think LN-penalty is the main problem here, as the most powerful form of the attack starts by filling channels with as many expiring HTLCs as possible, which is a problem for any channel design.  The only solutions I'm aware of are (1) temporary dynamic block size increases, which only increase the cost of the attack by a small multiple, (2) some form of time stop, although that increases the risk of capital losses from the time value of money, and (3) various bond designs, although they upfront accept losses to the time value of money.

Of those options, I think bonds may be under explored but I also think that the main downsides of time stop may be almost entirely mitigated by John Law's hierarchical channel factories design, which would involve channels being opened by three parties, with two preferred partners being able to continue exchanging funds in the channel even if the third counterparty initiated a force close that was taking forever due to a time stop. 

 You are pointing to the LN penalty as a problem but isn't it exactly what makes it harder for LSPs to pull forced expiration spam attacks? 

 The attack you describe was called "forced expiration spam" in the original LN paper.  I usually call it an "expiration flood".  Other people have other names for it.  See https://bitcoinops.org/en/topics/expiration-floods/ 

 running a routing node as a Tor hidden service introduces 6 hops between the client and service. the increased latency, timeouts, and failure rates associated with just these 6 hops has been enough for routing nodes to forego allocating liquidity and serving clients over Tor

if the average number of routing node hops per lightning payment is something like 3-6, and if all routing nodes in the chain were to run as a hidden service instead of clearnet (which i think Leo is proposing here), then all of a sudden that's 18-36 hops, which would dramatically increase latency, timeouts, and failed requests

that said, just because there's an ungodly amount of hops doesn't mean it's not feasible. there's just no commercial incentive today since the market is so niche and operating an LSP - even over clearnet - hasn't been a particularly lucrative endeavor to date

congestion among the Tor network is also a factor, but incentivizing folks to run more Tor nodes would introduce a new set of regulatory risks that doesn't exist today, and i'm dubious that there exists a strong enough commercial incentive in the near-term