>>585
Nodes broadcast "I'm interested in this board" and others interested in the same board scan around to find it (if I understand the DHT mechanisms correctly).
When you post, your node's IPNS reference updates; others pick it up, merge the difference into their copy of the board, and broadcast the new board; this repeats ad infinitum until every node sees the same board. It'll be one giant self-updating graph of replies pointing to posts pointing to threads.
There would be no way to trust who posted when, so every post must be a reply to another, and ordering of sibling replies would remain unspecified. Also, spam mitigation would probably have to involve blacklisting nodes.
You could run a node yourself or visit a website running the same software, and it wouldn't matter.

>>588-590,594
I don't think birthday collisions are an interesting problem on this scale or for this system, because you'd need a universe's worth of data for one to occur. Targeted attacks are more of an issue, but not in the near future. SHA-1 is unsafe, but the current cost to generate a collision is around $70,000, and the hash can be changed when SHA-256 suffers the same fate, as >>598 says. (Git, meanwhile, isn't interested in any of this because it is not designed for security. Trusting a commit is for the repo owner to worry about.)