> Right, which is why I said it would have to periodically scan the registered clients to ensure a minimum number of clients has each block to ensure redundancy.

That's the easy problem, not the hard one I was referring to: doing what IA does requires you to be able to crawl web resources and identify everything which needs to be available for a page snapshot to be usable. IPFS only helps with that in the sense that you can tell whether you have the same URL payload without requesting it — you still need to handle dynamic behaviour and that's most of the work.

> I think a slow, cheap but reliable archive is better than "more expensive but lower latency", so I'm not particularly concerned with timeliness.

What I would be concerned with is “more expensive, higher latency, and greater risk of irrecoverable failure”. Relying on volunteers means that you need far more copies because nobody has a commitment to provide resources or even tell you if they decide to stop (“ooops, out of space. Let me clear some up — someone else must have this…”), and the network capacity isn't just a factor for user experience — although that can prevent adoption if it's too slow — but more importantly because it needs to be available enough to rebuild missing nodes before other ones also disappear.