Henning:
Indeed, the proposal follows the description in p57-58 of
http://www.open-std.org/jtc1/sc22/wg14/www/docs/n1256.pdf, which dates back to 2007. (Some Haskell related deviations do exist, like dropping the final suffix, since Haskell doesn't need it; and requiring digits both before and after the dot.)
I think of the format as precisely representing the value "mantissa x 2^exp"; where the mantissa is written in hexadecimal, and the exponent is left as a regular decimal integer. The discrepancy is rather weird, but I guess it made more sense when the standard was drafted. More importantly, all the other languages (C, Java, Python:
http://www.exploringbinary.com/hexadecimal-floating-point-constants/) follow this convention as well; so it would be unfortunate if Haskell diverged.
For the change in semantics for "reads:" That is indeed unfortunate since we lose backwards compatibility. But it's a very minor one and I would be curious if anyone depended on the existing semantics for any legitimate reason. I personally do not see any issues with it.
-Levent.