This may be dumb question / suggestion but......
We know the aligner struggles with side on faces and to a lesser degree faces from above and below (depending on the angle).
Is there or could there be way to either:
-Highlight these faces (or any misaligned fr that matter) in manual mode and then reprocess / re-extract only these highlighted / misaligned faces in a more RAM / GPU /AI intensive or slower manner to improve the alignments before a final manual realignment?
-Highlight cerain faces as "side on", "from below" or "from above" in manual mode and again reprocess, but the AI knows to use an adjusted landmark template, for each of these highlighed angles?
I appriciate this is likley not a simple task, but you never know