Because even using a light buffer with R16G16B16A16 format if the roughness of the material is close to 0 the radiance of the pixel goes well over the ~65000 limit of the render target format.
Yes with FP16 render targets, you have to clamp into the usable range, or some GPUs will convert out of range values into infinity (because FP16 can represent Infinity and NaN)... And then when you try to blur that for bloom, you get big squares of infinity (or big squares of NaN). Your tonemapper will likely turn these into big squares of white (or black for NaN)
if the roughness is 0, your surface should be a mirror, the brightest spot in a mirror should be the intensity of the light source, you cannot get a brighter spot. so if your light source is < 65504, it should be fine.
The mirror BRDF is an infinite peak in the reflection direction and zero everywhere else. With an input brightness of 1, this generates a infinite result (although technically it's infinite energy density, not infinite energy).
When dealing with point light sources (also a delta function), and the common approximations where we don't measure the solid angle of our viewport's pixels or our light sources, a mirror returns infinite intensity, due to us not actually integrating anything anywhere.
If using physically sensible light sources and less-approximated versions of the rendering equation, then you're ok