Yes, there indeed seems to be a bug on very large a values, in L294 (and also L184). The algorithm does detect that the values are large enough for the result to be 1.0 (or 0.0) within the required precision, but then returns the wrong value. The threshold for the shortcut is hit on your second example, but not on the first one.
The current version v2.5 is affected as well.
The current version v2.5 is affected as well.