Lowering variance – but not removing it entirely – would probably be the best solution here. Without any indication of the HP value rolled, it's a bit hard to gauge the threat level of most early-to-midgame monsters beyond a rough idea. I can say "okay, this is an orc priest, and smite is dangerous." But I currently can't fully answer "can I kill it before it kills me; do I at least have a good chance of that happening?" The balance between the challenge that the monster represents vs. the potential reward from killing it is currently not as clear as it could be, and having it vary significantly from monster to monster of the same species is rather unintuitive.
Looking at the philosophy in the manual:
Major design goals
challenging and random gameplay, with skill making a real difference
meaningful decisions (no no-brainers)
avoidance of grinding (no scumming)
gameplay supporting painless interface and newbie support
Minor design goals
clarity (playability without need for spoilers)
internal consistency
replayability (using branches, species, playing styles and gods)
proper use of out of depth monsters
Large variance in enemy HP certainly increases the randomness of the gameplay. Whether it increases the challenge really depends on player stress, and I personally don't think there's much skill involved in dying to something with far more HP than you would reasonably expect it to have, given other monsters of the same race's appearance beforehand.
However, removing monster HP variance entirely means that there is no longer the same sense of relief and or heightened tension when a fight is going more or less easily than expected. It also removes the element of skill involved in disengaging from the confrontation, at least to some degree. Moderate or visisble variance solves both of these issues: enemy threat levels are more clear, and players can consciously decide, with a greater degree of knowledge, whether to engage or not, yet there is also the risk that a fight will go south quickly.
I don't think a relatively opaque mechanism behind enemy health generation has anything directly to do with the meaningfulness of player decision making.
Grinding is irrelevant to this topic.
Large variance in HP is both not noted in the interface and is something that new players are rather unlikely to figure out on their own, because combat from the player-damage-dealing side is also randomized. An alternative to reducing variance is, as suggested earlier, to highlight cases of extreme variance, and thereby make it clear to players that the creatures are different than the usual.
Large variance harms clarity, and the randomness on the player end (and the non-numerical visualization of damage dealt) means that there is never really any indication whether the player just hit weakly (as would be assumed) or that the monster has a greater amount of health than usual.
It's obviously internally inconsistent, as uniques exist. That's not really an issue here, since uniques are meant to be unique fights against known but powerful enemies.
Replayability seems like it would be provided for by the procedural generation, random vault-placement, random enemy spawning, random item spawning, random damage dealing, etc. I can't imagine why a more consistent (though still somewhat fluid) health value for monsters would detract markedly from replayability.
This has nothing to do with out-of-depth monsters.
Essentially, the big issue here is clarity and transparency. Two solutions that have been proposed are a prefix indicating a rough degree of the variation, or a tightening of HP values. I'd prefer the latter, mostly because I like being able to fairly easily estimate my chances against a given enemy, but those options both solve the issue here.