Why can't it? Based on historical performance we know the context independent value of a single through linear weighting. With observational data, we know how many extra hits or bases a fielder gives up relative to average. Assign the linear weight value to the marginal outcomes, and voila, you have a very good idea of relative performance.
Remember, the runs created formulas, be it RC or EqR or etc, have very high correlations to runs scored. The Linear Weights model has the highest correlation. The input variables are the offensive counting stats (1B, 2B, 3B, HR, SB, CS, GIDP, etc) and the output is runs. With each season, you get 30 equations that look like this....
1B + 2B + 3B + HR + BB + HBP + SB + CS + K + OUT + GIDP + SF + SH = R
...one for each team. With 13 variables and 30 equations, you can find the linear value of each variable. To get even closer values, you can average the results over a period of years. So, while you are right, it is just an approximation, it's got a pretty small error factor due to the multitude of historical data that can be used to create the model.
EDIT: Where I think the models fail is in the adjustments for Fenway. JHB noted earlier that Manny is just as bad by these metrics on the road as he is at Fenway. That, to me, suggests there is a problem with the models because of two things. One, his problem is limited range, which is greatly reduced at Fenway. Two, everyone acknowledges he is very adept at playing the monster. His numbers should be better at Fenway. I trust his road numbers for what they are because those parks are normally sized and have normal wall types (for the most part).