Lightweight face matching with OpenCV YuNet + SFace, no dlib

For checking whether a known person appears across ~100 photos, OpenCV built-ins cv2.FaceDetectorYN (YuNet onnx, ~230KB) + cv2.FaceRecognizerSF (SFace onnx, ~38MB) from opencv_zoo work with just opencv-python, avoiding face_recognition/dlib installs. Extract reference embeddings from a collage, take max cosine vs refs; >=0.40 is a confident match, 0.30-0.40 needs eyeballing. Misses cluster on upside-down faces, tiny faces, old scanned photos (age gap), and sideways phone videos (rotate frames per metadata before detecting). Build a numbered contact sheet of the uncertain ones and view it rather than trusting the threshold.