I believe you are omitting the fact that in the interim analysis the p value will be heavily influenced by the 180 or so patients who are much earlier in their course where the curve separation is much less. The time points of these alive patients would be censored but still contribute detrimentally I believe to the p value evaluation.
In essence I think if you assume that the populations for the 9902B are identical to the combined 9901,9902A group then perhaps you could simulate the log-rank result by plotting the combined 9901/9902A survival curve and superimposing the results of the first half of the curve on the overall result.
How you can accurately simulate the cox-regression model from the limited info we have is beyond me. Clark has suggested some modelling in earlier posts I am catching up on, but I do not know how accurate these are and how much they rely on a priori assumptions that may not hold true.
News 
Market Data 
Discover 
