JSTOR does exactly this with scholarly journals, and it works out pretty well. Recent issues are accessible only to paying customers.

Back issues (usually at least a few years old) are available via JSTOR for free in small amounts and through subscriptions for bulk users. I'm sure there's some reason to fight about the details, but from a distance it looks like a pretty good compromise.

Agreed. IA should take snapshots of the articles over time and then make them publicly available X months/years later. There's no reason to immediately publicly mirror the articles beyond people trying to get around paywalls.