I have decided not to release KiFinder as a signed app, whether free or paid. Privacy drives that decision. The preschool albums I use it with contain hundreds of photos of children from many families. I can judge for myself whether to run face matching locally on my Mac to find my own daughter. Packaging the same capability for any album would require a different decision, and I do not feel comfortable making it.
The ArcFace model also sets a condition for any broader release. Its weights were trained on MS-Celeb-1M, which was withdrawn because it was built from people's faces without their consent. I can use that model for a personal tool, but I do not want to distribute it. An app release would first require finding or training a model on consented data. It would also mean taking on the review and support work for a Mac app.
I still want other engineers to inspect the approach. I plan to publish the source under GPLv3 instead of distributing an app. People can read it or build it for themselves, while derivative work will have to remain open.