I think you’re correct, and this is where lumen’s ‘1 bounce+fallback’ design paradigm seems to be very smart. Most of the valuable visual behavior is happening on the first bounce of the refraction/reflection/GI, but you still need something to stand in for the other bounces to account for energy loss. Using the translucency radiance probes for reflections after the first bounces makes a lot of sense, because that’s normally obscured enough that you wouldn’t even notice at first.
I think you’re right to a large extent on needing refraction for caustics. However, there could be a case where one could use a simplified transmission model to handle directional lighting, and receive simple colored shadows as opposed to full-on reflective or refractive caustics. This has the added benefit of being more coherent than arbitrary ray-tracing depending on the technique, but no doubt still a perf. hit.