Just for fun, and to determine for myself what the base health and scaling was for warriors and pallies, I collected a small sample of players I knew in-game from the armory, recording paperdoll stamina and total health. I removed 275 health from those with the chest enchant, since that amount is not reflected in the stamina number, and tacked on unmodified at the end.
These are the data points I collected --
Paladin: (3905,45804); (3417,40924); (2736,34114); (3756,44314)
Warrior: (3732,45261); (3304,40981); (3780,45741); (3842,46361)
I ran these through a least-squares program to avoid doing a lot of arithmetic (and to easily confirm whether or not the data was exactly linear), where y is health and x is stamina. Also keep in mind that when you look at the formulae, talents are not reflected in the slopes, because the percentage effects of talents are already reflected in the paperdoll stamina number itself.
Both regressions resulted in a correlation coefficient of exactly 1, which means that the data is perfectly linear.
Paladin: y = 10x + 6754
Warrior: y = 9.9x + 8278.64
Now the point of posting here, getting a slope of 9.9 instead of 10 exactly looks very suspect in the warrior regression.
This isn't possibly the result of any kind of statistical error in the data set, because the line is not approximately fitting the data, the correlation is perfectly linear.
Also the size of the difference feels too large to be the result of any kind of roundoff error. For example, if a warrior tank has approximately 3000 stamina, the difference between a slope of 9.9 and 10 is comparable to an entire stamina gem (not counting talents).
EDIT: This is all for naught! There's an error in the warrior formula I originally posted. The real formula is -
y = 10x + 7941
This does however explain why the health pools always end in 4's for pallies and 1's for warriors.