Very high-dimensional space is unintuitive. It can help to imagine a cube as N increases - 2^N corners and vanishingly small volume to surface area.
This gives useful intuitions e.g. cosine distance ignores length and only looks at angles, but this matters less with higher dims; by contrast the euclidian length becomes less useful.
I get that cosine distance is easier to calculate, but isn't sorting by cosine distance the same thing as sorting by L2 norm, assuming that the vectors are already L2 normalized?
Do I understand you correctly, that you say cosine distance is a better measurement of similarity than euclidian distance when there are many dimensions?
This gives useful intuitions e.g. cosine distance ignores length and only looks at angles, but this matters less with higher dims; by contrast the euclidian length becomes less useful.