I too am moving towards a flat hierarchy with dates after attempting Zettlekasten and PARA. I like your "Zettlekasten layer on top of the fs" idea. Def going to do something similar.
- scan to an image/as many as needed
- mogrify -deskew 90%
- for i in .png; do; convert $i ${i/ng/df}; rm $i; done
- pdftk cat output aname.pdf
- ocrmypdf --force-ocr -c -i (wrap tesseract and others)
I normally get a good enough scan. If it's not the case I might manually edit images than feed to the rest of the pipeline. Gimp G'Mic QT Repair Scanned Document filter works IME far better than unpaper and others, but it's not that automatic, need to be tuned case by case. However that's rare.
For notes, personally I go for time-based notes, similar to ZK but with the time-constraint, I've described the system in this page.