Re: New paper: Accelerating Bitcoin's Trasaction Processing

2) This is a local-only change that does not require consensus. It's okay for light nodes to still follow the most-work chain. What this change does is provide miners specifically with a higher chance of ending up on the final chain, by better estimating which fork has the most hash power behind it.

Not sure why you call it "local-only change", because as we are discussing here, it appears that there's a risk for netsplits that don't re-converge, especially if different nodes will follow different rules.

There is no possibility for non-convergence. The most-work tree eventually wins out. Always.

There could be an extreme scenario where two fractions of the network work on different trees, and each fraction uses anti-DoS rule that rejects blocks from the tree of the other fraction unless they arrive in a batch that proves that the PoW of the competing tree wins, but because of propagation lag each fraction solves more PoW blocks in its own tree so by the time the batch arrives from the other tree, that batch is already losing and therefore rejected. No?