Remember that if you estimate it by looking at the combined log rank curves, you are not taking into account that the combined dataset was imbalanced against the treated group (which is why Cox HR was so much better).
Just an FYI - even 9901 exhibits a much lower HR in the first half of the curve (around 1.2 to 1.3) than in the latter half (2.0 to 2.2). It isn't just about the imbalance in the 9902a (which had an HR of 1.0 in the first half of the curve).
As for more comments, I'll defer until I have gotten to my model and figured out more details about what assumptions I would settle on.
News 
Market Data 
Discover 
AI
