If the existing code was using int32, a switch to anything larger would cause integer overflows or JSON parsing errors in languages with strongly-typed fixed-width integer types.

Any call from a typed langage distinguishing between 32b and 64b integers (that being most popular typed languages I reckon) would break if it had assumed / used the smaller of the two.

TBF using the negative range could also break callers distinguishing between signed and unsigned if they’d used the latter on their side depending how the API was documented.