To be perfectly honest I haven't spent much time on team totals but I think the real problem here is that you can't overlook the correlation effect (to which you alluded) without "breaking" these calculations. In the NBA there is
substantial positive correlation between game opponents' results versus their imputed team totals -- a correlation coefficient in the neighborhood of about 40%.
Technically speaking, however, the issue isn't just one of correlation but rather of the lack of independence because team scores (even if deviations were uncorrelated they still might not be independent).
So even though the probability of Team B scoring > 96 points may indeed be 50%, the probability of Team B scoring > 96 points conditioned on Team A scoring (say) 106 points could very well be different than 50%. As such it's entirely likely that using the HPC as you've described will produce biased estimates.