You've hit the nail on the head with your question about "why not allow any narrowing at all, as long as it's within range?". The answer is, in two words: type safety. When you allow any narrowing at all, as long as it doesn't change the value, you've still broken type safety. Though you haven't changed the *value* of the variable, you've changed the possible *range* of values that variable can take on, and that can have subtle and surprising effects depending on the usage.
For example, if I write a method:
This method converts floats into extremes depending on whether they're less than 1 or not. Now if I could call this method with a double, and it returns a float so I can assign that return value to a double with no problem, look at this code:
Most reasonable people would expect dExtreme to be either Double.MIN_VALUE or Double.MAX_VALUE.
This is a pretty weak example because you can just look at the API for the extremify() method and sort it out, but if this procedure is happening in the middle of a method instead of with a call to an explicit, documented method, you could get thrown. Also, there are other cases for which I cannot conjure up examples at the moment that would lead to ambiguities in the language.
Overall, the general rule should be absolutely no implicit narrowing period, and it would be fine with me if Java didn't have any exceptions to this rule. The exceptions they do allow have to do with very specific conversions for which no ambiguities can arise and for types such as byte that are truly "flat"...they are just strings of 0s and 1s and there is no connotations associated with the data...it is truly just "raw data", and when you can't make any assumptions about type, it frees you from certain restrictions.
I know this is all very vague, but hopefully it gets you close enough to the path you can think it through and figure out the reasons from here on. I would, but it's been years since I dug into these issues and it's just not coming to as I sit and type.
sev