It may be because the lambda parameter in the soft-core-softed potentials has a more complicated functional form and so your analytical derivative df(lambda)/dlambda is not 1.0
Mixing lj/cut/soft with compute ti is not simple…
compute fep was made to work with the core-softer potentials, so it should work fine.
Maybe you can test compute ti against the other methods on a free energy route that does not involve starting from 0 interactions and in this case you don’t need the soft potentials.
Hope this helps,