In the documentation the phrase "nearest integer" is ambiguous for half-integers.

In addition, this function rounds half integers to plus and minus infinity (introducing a statistical bias when negative and positive values are not equally likely to occur). The most common way of rounding half-integers is to the nearest even integer (e.g., IEEE 754).

Funnily, you can round infinity!

?- X is round(-1.0Inf).
X = -1.0Inf.

?- X is round(+1.0Inf).
X = 1.0Inf.

And even NaN

?- X is round(nan).
X = 1.5NaN.