You (and other bot operators) can do some experiments to determine how required time to find solution depends on target time, then we can construct some function that is good approximation of this dependency and use it instead of 1/t.
The more I am looking at and thinking about the "cumulative work warp" attack that you've brought up the more I am realizing that this will likely be a necessary step in any resolution for it.
I'll be working on establishing an approximation.