But altruistic "voting your conscience" also redistributes the rewards, just like downvoting does.

The difference is that this has a huge opportunity cost compared to voting to maximize your own reward (either directly or via some scheme). Since the cost is very high, it is unlikely to see such altruism persist. Sure, not everyone will maximize their own reward, but when the incentives are misaligned in this manner you expect the economy to shift more and more in that direction over time, which is exactly what has happened. I don't believe this has anything to do with stake distribution or the number of voters or any such thing. It is more that it is baked into economic design.

By contrast, once the high cost of downvotes themselves is removed, good downvoting vs. no downvoting vs. bad downvoting does not have a high opportunity cost (in effect virtually none), so it is far more likely that altrusim will succeed. If people had to pay money (or forgo a significant payment) every time they edited wikipedia, hardly anyone would ever do it (in fact, those who did would be more likely to be promoting some sort of manipulative agenda and often damaging the end product). But because editing wikipedia is very cheap/free and frictionless, you do get altruism on a large scale. Crowdsourcing works, but not if you expect people to pay a lot to participate. The only way to get this with Steem is by crowdsourcing downvotes.

This not just a question of downvoting being used actively to 'fight abuse', which is the traditional way we have looked at it. It is a profoundly different setup of the incentives, which can't be accomplished by upvoting, because the economic 'pull' of voting to direct rewards toward oneself and/or one or more collaborators is always there with upvotes. With downvotes it is not.

The only way I could see it happening would be the rise of a mutual-voting group

Such groups would certainly arise if there were significant money at stake, probably cleverly packaged into a game, challenge, service, meme, corporation, investment fund, etc. That is the nature of human creativity.

In any case, we can't simply wish ourselves into a system with evenly distributed stake and billions of users either, even if that would somehow manage to work. We have to need to have economic rules which are robust to real world conditions. Which means downvotes.

If a sufficient amount of stake becomes involved in such arrangements, then the reward system by itself no longer incentivizes upvoting based on content.

Well, no. A superlinear reward system never, in and of itself, incentivizes upvoting based on content. Structurally it only incentivizes concentration. And concentration that rewards the voter is incentivized most of all. Content is actually irrelevant unless there is a large enough portion of the stake voting on an altrustic basis, and as I've argued above, it is very unlikely to see altruistic upvoters persist on a scaleable, sustained basis. (Things might start out that way, but they will surely devolve over time.) The only real hope we have is altruistic downvoters.

Please give this some more thought. It took me the better part of two years to finally reach the conclusion that downvotes are essential (and not just theoretically-possible expensive downvotes that are hardly ever used, the way we have now), but I'm now quite confident it is correct.