This seems quite interesting, do you know the advantages of this method over a simple 1x1 downsample luminance average?
With a simple average a few very bright HDR pixels (sun, fire or weapon muzzle flash) can change average luminance and over darken the entire image, so it's a good idea to skip a few % of brightest pixels. Similarly if you don't ignore most dark areas then entire image can be too bright. See "Post Processing in the Orange Box" for an explanation with some images.
Also I still having problems to understand how to calculate exposure from luminance if the tone mapping operator does not take in count the exposure.
Exposure = some_constant / averageLuminance and then multiply every pixel by exposure just before the tonemapping operator.
To be physically plausible, Bloom/Glare/Glow should added before tonemapping since it simulates scattering due to diffraction effects in the eye/camera lens, which are wavelength but not intensity dependent.
Yes, You are perfectly right.