It seems like the debate has become intractable because in addition to the base disagreement about the effectiveness of filters, there is also a debate about how to decide on controversial changes to the reference implementation. Core does not seem willing to merge a PR they disagree with and Knots does not seem solely content with building in the Knots repo. So polarization happened because both sides feel like they have "ownership" (probably not the right word) of the reference implementation.

I appreciated @flaco's take on this from yesterday (#1234973), and would be very interested to read your thoughts in response.

This feels like one of those high-density statements the unpacking of which leads to a lot of learning on my part.

I'll spend more time on it, but from what you said and the attached mailing list thread,

Is it that everyone's watchtower needs to know everything, or that everyone's watchtower needs to know miner's mempools? In the thread Todd says that the double spent txs had "were highly relayable/minable" but that the problem was that Shapeshift relied on something that used percent of nodes to which a tx has propagated as a proxy for confidence. This makes me think that the entire relay network can be the same, but if miners are willing to mine nonstandard, zero conf won't be reliable.

But all that aside, I'm still trying to figure why LN is a zero conf protocol: in the case of uncooperative channel close, first one to get a tx in a block wins. But what about penalty transactions? Don't they mitigate this?

Or is LN like a zero conf protocol because I don't really have my sats till I return to the chain?

I think I still don't have this fully in my mind, but I don't mean to take up your time with it (more just writing it out for the practice -- I'll spend a while on it and see where I end up).