Writing a generic generator that can also take sensor input would remove cumbersome dependencies and make the MTC frame work more flexible and easy to use. I can start here.

The approach and lift directions depend on the grasp candidate which is related to IK feasibility. Is there a way to propagate the pre-grasp and post-grasp poses to other stages after the IK stage?

In principle, you can assign arbitrary properties, e.g. pre- and post-grasp hand postures, to any solution. And you can ask MTC to forward these properties in adjacent stages (setForwardedProperties()). I never finished to write an example pick and place task relying on these concepts, instead of hard-coding them into the task program.
Maybe, it is a good starting point to rewrite the existing Pick and Place containers to use this more generic scheme. This will be needed anyway if you want to pass such properties from the grasp generator stage.

I started looking into these containers. It looks like in order for them to properly pick/place both containers would need to modify the planning scene to control which collisions are allowed during grasping, lifting, and placing.