Yes, IEEE 1584 is a curve fit exercise. There are problems at the end points but possibly not what you described. See here:
http://ep-us.mersen.com/resources/media/articles/WP-Improved-Method-ArcFlash-Analysis.pdf The "bogus fit" problem shows up on the high end and sometimes on the low end but you can easily sanity check when this happens (Iarc > Ibf).
As to your example, this happens in the real world quite often. As you lower the available fault current, the inverse time curve hands you a great deal of pain. If the upstream protection is a fuse (you did not specify), it likely trips in 1/4 cycle. Let's just say for argument's sake that it's 10 kA available fault current and assume that incident energy is linearly related to current (it's not by just saying). Now you are reducing the bolted fault current to 1050 A, and the arcing fault current is 85% of that or around 850 A. This might be beneath the instantaneous trip of the 100 A circuit breaker in which case 2 seconds is not unreasonable, or 120 cycles. So we've reduced the available fault current by roughly 10 fold but also increased the trip time by nearly 500 times. So some more napkin math suggests that the incident energy is about 50 times. 0.3 x 50 = 15 cal/cm^2 so the 6 cal/cm^2 value doesn't seem so unreasonable any more.
This is one of the little lessons to learn about arc flash...if at all possible, try to set your protective devices so that the arcing fault current is within the device's instantaneous or at least short time trip curve. Otherwise, be prepared to deal with some very large values.
As to the example of using a wire as a fuse, again, not an uncommon scenario. The jumper to medium/high voltage surge protectors is quite often sized intentionally to be used as a "copper fuse". This does not show up in the model either. Sometimes you have to use a little thought when applying IEEE 1584.