So, everything was clear from the specification point of view, but when I moved to reading through some of the original implementations and reimplementations, I ran into this:
f = ldexp(1.0,rgbe-(int)(128+8));
This piece calculates the signed exponent (<128 are, by specification, defined to be negative) and applies it to 2.0f. Why does he add 8 after the fact? And why doesn't he just write 136? That's biasing the encoded exponent information in favor of the smaller values, right? But why? All the values I load with this seem to be in a really cramped floating point space [I'm not even sure anything uses the space above 1.0f]. I understand the ratios remain consistent, but I kind of prefer to rely on original measured radiances.
http://www.graphics.cornell.edu/~bjw/rgbe.html — The code here states that is the standard way to decode "real pixels", which has no basis in the specification and no rationale offered.