Pairing for knowledge raises the bus number from one to two, which doesn’t help for that long in the current market for devs. Documentation is what endures.
I worry you missed the "documenting tacit knowledge is extremely difficult" part. Tacit knowledge is notoriously hard to document in a way that successfully transfers the knowledge from one person to another. Pair programming is at least an order of magnitude more effective, if not more.
If two people build a feature, two people are slowly forgetting how and why until they move on, and nobody else gets it at all. It is easier than documentation (which I agree is hard enough to more than double the dev time), but it only postponed the problem a little without solving it.
