Establish a production-grade, reusable dora-rs C++ node for ZED cameras that supports native object detection, open-vocabulary inference, and optimized point cloud routing.
Reduce dora-zed to a camera source: BGR image, depth map and intrinsics. Detection moves to its own repo and the ONNX Runtime dependency is dropped. Add the depth output that downstream perception requires, match the depth configuration of the reference camera driver, fix camera recovery so a dropout cannot block the dora event loop, and build against dora 1.0.
Exit criteria: a dataflow consuming image_bgr, depth and camera_info runs against a real ZED; unplugging and replugging the camera recovers without restarting the dataflow; the node builds on Jetson with no ONNX Runtime installed; dora validate reports wiring and type annotations OK.