Mixed results might be best, tight surfel GI shadowing the heightfield GI, even if the shadowing distance is small it will look better. As for thin geometry and particles, I’ve found success with what I call “box sampling”. Just take the AABB of your asset, or whatever box you want, and sample GI in a sphere from the center. You can then construct, whatever, a spherical harmonic term to apply to the resulting geometry and you’re still going deferred. Works with transparent stuff too but then you don’t really get specular for glass or whatever unless you sample enough to not make it worth it to begin with. Still, you can specular sample as normal for a pane of glass if you have that set up, then “box” sample for the diffuse, if it’s transparent enough it looks fine.
There’s errors from this of course. But you can sample higher from a single spherical sample than trying to go per pixel on all that stuff, and then apply whatever hacky corrections you can think of like blending it with nearby “boxes” or etc.