cv2.imread and VideoCapture fail silently — what are the limits of making OpenCV loading reproducible across machines?

Goal

I want a repeatable Python development environment where image and video loading behaves identically on every developer machine and in CI. The sticking point is OpenCV's documented silent-failure semantics: cv2.imread returns None instead of raising when a file is missing, unreadable, or uses a codec not compiled into the build, and cv2.VideoCapture only reports failure through isOpened().

Constraints and uncertainty

Decoding support depends on the specific wheel or system build (opencv-python vs opencv-contrib-python vs headless variants), and the video backend (FFmpeg, GStreamer, V4L2) is selected at runtime per platform. I know I can pin the wheel version, inspect cv2.getBuildInformation(), and pass an explicit API preference such as cv2.CAP_FFMPEG, but OpenCV offers no strict mode that turns these silent failures into exceptions.

Questions

• Is wrapping every cv2.imread call with an explicit None check the accepted limit of what OpenCV guarantees, or is there any supported strict/error-raising mode?
• Does pinning an exact wheel version plus an explicit capture backend flag fully eliminate cross-machine differences, or can codec availability still vary within the same wheel?
• Are there documented cases where OPENCV_VIDEOIO_PRIORITY_* variables are ignored, making backend pinning unreliable?