I am guessing that the reason most regress toward the mean is because they found that produces the best results. These aren’t about producing a result that mirrors the eventual ERA distribution. They are about finding the most likely result for each individual player. If you wanted the projections to look like the eventual ERA leaderboard, it wouldn’t be hard. Create a curve based upon leaderboards from recent seasons, use the projections as a ranking and place the players along the curve in that order. The thing is, if that were more accurate on a player to player basis over a number of seasons, that is likely how it would be done. It is not.
You aren’t wrong, though, that the mean is far from the only relevant part of a projection system and that projecting reliever to starter is difficult on numbers alone.