That's a defect in the model - which is Frank's model, not mine. An expected outcome of 1.5 rounds should assume a limit function such that P(kill orc)=1.5 as the number of rounds goes to infinity. Typically you'd cut it off at some amount where the probability of reaching round N is vanishingly small, but assuming that a combatant eliminates his opponent on round 1 or round 2 and the combat never goes past that, ever, is more simple a model than is useful... especially in the context of standard deviation effects on the combat's progress, and how "failure" is being defined where death is a-okay.
It's actually a complicated series of probabilities, where, say our PC kills Enemy on rd1 50% of the time, rd2 40% of the time, rd3 8% of the time, and rds4+ 2% of the time... that gives an expected number of rounds of (.5)(1)+(.4)(2)+(.08)(3)+(.02)(4) = 1.64 rounds of expended resources per PC.
Then for the two-PC/two-Enemy problem you'd have a matrix of each interlocking probability, such that the combat ending on rd1 is (.25)(2), rd1.5 with one dead orc in rd1 as (.2)(3)*2, rd2 as (.16)(4), rd2 and rd3 as (.032)(5)*2, rd1 and rd3 as (.04)(4)*2, rd3+ as (.0064)(6), with an expected outcome of... 3.0184 rounds of expended resources, or 1.5092 resources expended per PC.
Huh.
So that's pretty much showing that per-group is better than per-PC for encounter generation, at least as a first step.
I'd still like to see that with the option for there to be PC death, though, and the orcs manage to focus fire on a lone PC.
...
That does still leave the standard deviation issue outstanding - in a game where death isn't a minor speedbump, accumulation of those lost resources into a single character is going to have an increased marginal effect on lost resources.
If, for example, we take the two-PC/two-Enemy problem, if there's a 10% of PC death (just to pick a number), if resurrection costs are, say, 10 units of in-battle resources, then those marginal cases are going to heavily adjust the expected resource consumption overall. Median cases won't be affected, but it'll throw off the mean.