The effect of the flag nlo_mixed_expansion

In v2, the syntax
generate p p > t t~ [QCD]
*strictly speaking* did not mean "Do NLO QCD for ttbar production". Rather, it means "compute QCD-type loops and real emission for ttbar production". The difference between the two is subtle; indeed, for most processes (ttbar is one of them) there is actually no difference.
However, take VBF. in v2, loops that involve non-colored particles (aka the pentagons) are always discarded, because they are not considered as QCD loops. However, they are a (very small) NLO correction at O(alpha_s) relative to the Born. Hence, they are now included in v3. Single-top is another example (keep in mind that typically, but not always, the discarded contributions are either very small or zero e.g. due to color conservation).
While the v3 behaviour is the most consistent, it is however not always the most convenient practical. Also, one may (rightfully) want to have the same behaviour in the new version.
In general, processes for which there is an effect are those with t-channel vector bosons at the LO (VBF/VBS/single-top/etc)