My strategy does not change other voters' strategies. Secret ballots prevent this type of coordination. That's my point. The collective strategy of 1/3 of the electorate does change which two candidates are most likely to win, but my individual strategy does not meaningfully do that.
> they each have the selfish incentive to give their second choice a higher score than their third choice to prevent the worst-case outcome.
If a voter values preventing the worst case over achieving the best case, then the optimal strategy is to assign maximum scores to every candidate except the worst case. Hedging by assigning a non-maximal score increases the chance of the worst case compared to that approach, in exactly the same way that it reduces the chance of that compared to assigning a minimal score.
I'll grant that my specific tactic is predicated on a preference for achieving the best outcome rather than avoiding the worst one, but the best tactic for someone who finds avoiding the worst-case to be more important also only requires extreme votes.
> A's voters prefer that B win over C and B's voters prefer that A win over C, so they each have the selfish incentive to give their second choice a higher score than their third choice to prevent the worst-case outcome.
Expressing that preference directly reduces the likelihood of each such voter's preferred outcome, even if a single voter does it. It affects the chance of the worst-case outcome only if voters on both sides of the A/B division do it. The secret ballot prevents any kind of enforced coordination. This is exactly a prisoner's dilemma.
> Requiring your preferences to be expressed statistically increases the error bars for no reason.
You don't know the exact score each candidate will end up with absent your vote -- if you did, you could analytically determine a single-vote strategy that gives the best available outcome. Since you don't know that, your choice of an intermediate score is a statistical expression. It's just expressed in terms of the uncertainty in what the rest of the electorate is doing, not in terms of a coin flip. It does not meaningfully increase the error bars (in a large election -- say, >1k voters) because the former uncertainty quickly dwarfs the latter.