Interesting question. There are subtle issues here which I haven't seen documented.
It seems INDIRECT("A"&ROW()) returns an array consisting of one element which is a reference to a cell - not the value in that cell. Many functions cannot resolve this type of data correctly but a few functions such as N and T can "dereference" the array and return the underlying value.
Take this case where there are two elements in the array:
=SUM(N(INDIRECT("A"&ROW(1:2))))
This returns A1+A2 when array entered but it only returns A1 when entered normally. However changing ROW(1:2) to {1;2} in this formula returns the correct result when entered normally. The equivalent SUMPRODUCT formula returns A1+A2 whether array entered or not.
This may be related to how the arguments are registered in the function. According to http://msdn.microsoft.com/en-us/library/bb687900.aspx there are essentially two methods to register function arguments to handle Excel data types:
Type R/U: "Values, arrays, and range references."
Type P/Q: "Excel converts single-cell references to simple values and multi-cell references to arrays when preparing these arguments."
SUM arguments seem to conform with type R/U while SUMPRODUCT arguments behave like type P/Q. Array-entering the SUM formula above forces the range reference argument in ROW to be evaluated as an array whereas this happens automatically with SUMPRODUCT.
Update
After a little more investigation, here's further evidence that might support this theory. Based on the link in the comment, the formula =SUM((A1,A2)) gives the same values as:
?executeexcel4macro("CALL(""Xlcall32"",""Excel4"",""2JRJR"",4,,1,(!R1C1,!R2C1))")
Registering the last argument as type P by changing 2JRJR to 2JRJP gives an error in this case but does allow for single area ranges like !R1C1:!R2C1. On the other hand, changing the 4 (xlfsum) to 228 (xlfsumproduct) only allows single area references either way it's called just like SUMPRODUCT.