Micciancio's notes are correct, and the standard way of explaining things, so let's elaborate some.
First, its worth mentioning this has nothing to do with (Gap)SVP specifically, and everything to do with what are called promise problems more generally.
A promise problem is a certain relaxation of a standard decision problem. They are meant to relax the notion of correctness for the problem. The standard notion of correctness can be summarized "The algorithm is correct on all inputs". There are two natural relaxations of this
Randomized classes (like BPP, ZPP, (co)RP). On any particular instance, the algorithm now only needs to be right " on average", where you average over internal choices of random coins.
Promise problems, where you are fine with the algorithm making mistakes on "hard instances", but it always has to be right on "easy instances".
In particular, on hard instances, an algorithm can be completely incorrect, and we don't care. As long as it is right on easy instances, we say its right overall.
To bring things back to GapSVP, the hard instances are when $d\leq \lambda\leq \gamma d$. So we say an algorithm solves GapSVP if
Given an instance $(L, d)$ (a lattice, and distance bound) that is easy, meaning $\lambda_1(L)\leq d$ or $\lambda_1(L)\geq \gamma d$, the algorithm returns the right answer
Given a hard instance, the algorithm returns whatever it wants. We don't care.
In particular, given the same hard instance twice, an algorithm could return both answers (it doesn't have to be internally consistent). We don't care --- we only care how the algorithm does on "easy" instances, measured by $\gamma$.
