Sure although I think you're overstating that there are "no pathological cases" with clustering. Any conservative culling system will still be primarily bound on the pixel that has the most lights. Unless you actually start approximating illumination with the clusters or are willing to reschedule the entire workload (work stealing might work here, but limited somewhat by GPU hardware), you still have to evaluate every light that affects that pixel serially.And with thousands of lights, a single tile can get pretty expensive, especially given the parallelism needed to feed modern GPUs. With 3D clusters there are no pathological view points, and work size is closely related to light density.
But yes, I agree in general that clustering is more robust to worst case performance. That said, I'm not totally convinced that it would matter a lot in practice compared to something simpler.
Oh a shader compiler is always still free to just flatten the branch if it wants so there are no guarantees, but in practice I've found it to be fairly reliable in HLSL at least.I'm not aware, however, that you can control branching on that level in glsl, or CUDA (as you can in hlsl), so results might vary in practice.