fix: Python 3.10 datetime.fromisoformat fails on 5-digit microseconds from Supabase — pad to 6 digits before parsing

debug: add /debug/full-test endpoint to trace exact DB path per employee

fix: raise match threshold to 0.65 and return best_distance on not_registered

fix: hardcode CORS allow_origins=* to fix mobile connection errors

fix: remove silent exception swallow in load_encodings; surface debug count on not_registered

debug: add /debug/employees endpoint to diagnose clock not_registered issue

fix: use fresh Supabase client per request to avoid stale HTTP/2 connection errors

feat: scan all employees when no admin_id provided (remove no_admin gate)

fix: wrap all endpoints in try-except to expose actual error instead of 500

fix: convert PIL RGB to BGR before passing to InsightFace

Use insightface instead of face_recognition — matches Dockerfile

Add admin_id multi-tenancy to /register and /clock

Fix: pad Supabase timestamp fractional seconds to 6 digits for fromisoformat

Fix: robust timestamp parsing and expose 500 error details

Lower similarity threshold to 0.3 for better recognition

Fix: upgrade onnxruntime to 1.18.1 (supports NumPy 2.x)

Fix: pin numpy<2 for onnxruntime 1.17.1 compatibility

Fix: add build-essential for InsightFace Cython extension

Switch to InsightFace+ONNX — no compilation needed

Fix: use python:3.10 base image for clean environment

Fix: add face_recognition_models package

Fix: install numpy and pillow via conda to avoid compilation

Fix: use conda pre-built dlib to speed up build

Add FastAPI facial recognition backend

initial commit