I have a number of SOT23 parts and use triode with them. Every 6-10 parts it will reject one with seemingly no reason.
The black and white capture window looks identical in rejected and placed parts. I have tried

Turning down accuracy - I think I might be using 2 now but have tried three.
Playing with contrast and illumination settings
Increasing and decreasing the capture window

It would really be useful if the optical recognition would output some errors when it rejected parts - for future improvements.

Do you have any advice? Is there any debugging available at all in the optical recognition system?
Anyone else have any advice?

One idea that might work will would be a Neural Net with training that could be crowdsourced - this seems probably beyond your software effort and also more open than the software effort appears to be now.

I'll give it a try. Where does output appear?
Are there logs?

In recent projects for the SOT23, I have chosen the Vision Model "04: Footedge 2".
Problems with the installation of SOT23 have decreased significantly.