Not really answering your question (which is answered elsewhere in the thread), but these things are independent in an interesting way. Sybil attacks, in the simplest sense, depend on running consensus across a group where you don't have trustworthy membership information. If you do, then it turns into an authentication problem, which is different. Sybil attacks are an attack on the protocol that controls jury membership (or authenticates jury members), rather than the consensus protocol itself.
Upon re-reading the paper, I also noticed that it says "In the beginning of the Propose step the [deterministically chosen] designated proposer for that round broadcasts a proposal to its peers via gossip" but it's not obvious to me what happens if that proposer fails to participate. Is the algorithm allowed to skip to the next round? If so, is there some guarantee that the gaps can't later be filled in, and the history retroactively modified? (The DLS algorithm solves this problem by requiring all nodes to propose a value.)
You can prevote or precommit more than once per height (on different rounds), but a commit for a height must be the last signature for that height.
The sudden mass bonding is a problem for light clients doing a certain type of desirable SPV. We're considering options to mitigate this issue. One such option is to throttle the bonding and unbonding at the protocol level, or to introduce a second class of coin for validating and controlling issuance of it.
Section 6.5 purports to give a game-theoretic justification for why a node should include as many valid signatures from its peers when proposing a block, for fear of retaliation. But nothing stops me from dividing my voting share among, say, a thousand nodes which appear to be unrelated.
If one of those nodes decides to "defect" (steal a larger-than-intended share of fees by excluding some signatures) then the expected time before it is selected again can be made arbitrarily large, by making the node's voting share arbitrarily small. If that expected time is significantly larger than the unbonding period, then there's no way to retaliate, is there? It can just unbond and then transfer its coins to a "new" node.
EDIT: I originally typo'd "arbitrarily large" instead of "arbitrarily small" above, but hopefully the meaning was clear.
More bandwidth && less latency && more CPU helps.
1. You assume the initial distribution of tokens is trusted. If someone presents you a consensus without a history to make you believe that initial distribution isn't full of sybils, distrust.
2. Assuming a large trusted initial distribution, operationally you have to believe that 1/3rd of the validator nodes don't want to double spend/ present an alternate reality to you.
I don't see how your point 2 solves anything, though; it's just failing to acknowledge the problem. Why is it reasonable for me to think that a supermajority of the nodes are honest? If validators are supposed to be computationally cheap to run, compared to bitcoin miners, then what stops an attacker from running lots of them to get a larger voting share?
Personally, I agree with you.
I am writing the "validators contract" for tendermint right now. It will have this functionality.
If it does, I expect that it would reduce the likelihood of a Sybil-styled attack during the early days of the network -- with the risk of introducing a different vector, direct coercion of the node owners.
Is the plan to make the hard-coded genesis nodes as geographically diverse as possible?
So far I've only skimmed the white paper, but I didn't see any response to this problem. (If the answer is "good guys can make lots of fake nodes too" then you're creating a computational arms race that will inevitably lead to just as much wasted effort as proof-of-work.)
There's also the minor practical issue that every node has to do an amount of computation and network traffic that's linear in the total number of other nodes, meaning the total workload grows quadratically with the size of the network.