Keras did make it into tf.contrib now (with the release candidate for 1.1). I don't see it going away any time soon. TF-Slim though I'm not sure about. Never really used it myself.

That's the big long-term question. The nice thing of this is really just having Google backing.

Keras, though, still really incredible for usability.

Generally all of these high level interfaces are somewhat interoperable because of the graph nature of Tensorflow. It might make your code harder to read but you can use some layers from slim (for inception-like nets this is a decent reference https://github.com/tensorflow/tensorflow/tree/master/tensorf...), and some layers from Keras. I don't know as much about sonnet, but it looks like a similar story.

Not sure. I think Keras is more meant for education and off-the-shelf models, not so much for research due to its abstraction. Or in other words, its less customizable. I think the probably best trade-off between low-level code and coding productivity would be tf.layers, which is also in tf.core now.