So you're actually confirming my point? If I follow 1k users scattered across the network, my client either needs to:

a) connect to hundreds of different relays (massive overhead for the client)

b) hope those users consolidate onto a few large relays (centralization)

The "bottom up model" works fine at current scale, but saying "the global state problem is supposed to be centralizing" is admitting the architectural limitation I'm talking about. Its not possible to handle millions of users without either client-side complexity exploding or relying on centralized infrastructure.