You probably couldn't have an array that big and even if you could, it's likely this function would accept it at compile time. The problem here is that you're passing it an array that it is advertised to handle, yet it fails.
Make two arrays, and have the end of one act like a linked list to the start of the second. When a [x] value is larger than the size of the first, just subtract the first's size from x, and find that value in the second. It's even still O(1) time.
Viola, you've now got nearly double the storage of what you can reference, and it'll work in a generic case, but you'll never be able to refer to anything larger than your index-number can reach.
To make matters worse, do you know if your library isn't doing this for you? It'd allow you to extend your array without having to copy the whole thing, which saves a LOT of time on frequently-lengthening arrays. The extremely small cost of checking an int's size before actually accessing the sub-array necessary is often negligible compared to duplicating potentially billions of values.
This contrasts with the overflow defect which would cause the mergesort routine to fail when given an otherwise perfectly valid input.
I don't think you can do that in Java. In the more general case, you're right, but I think you can solve it by just using a long long or whatever your language has instead of an int for the appropriate calculations.
I also question the mergesort assertion. I've implemented mergesort several times, almost always expressed in terms of writing to/from pipes of data (that under the hood eventually wound up going to disk I/O in some way). Those solutions will not suffer from any form of this bug for the simple reason that you never access anything by index.
The curious might wonder why I felt a need to implement such a well-known algorithm myself. Well in addition to the times I did it for fun, one time I had a large dataset to process that had already broken the database, and the computer I had to process it with didn't have enough disk space for the whole dataset. So I needed to do a mergesort while keeping all data that hit disk in compressed form...