Calibration + Live Visualization

[spark032]

New Features

• Differentiable-render point-cloud tracker (new diffrender_tracker/ module) — aligns two Spots' point clouds via colored ICP with temporal warm-start and FPFH/RANSAC bootstrap; includes a gsplat (CUDA) and soft renderer, SE(3) pose fitting, and offline/real test harnesses.
• Live two-Spot ICP viewer (live_pair_icp.py) — streams both robots, fuses their front-camera clouds, and aligns them live in an Open3D window.
• Live fisheye point-cloud ingest (live_pointcloud.py, spot_ingest.py) — fisheye back-projection to colored clouds, voxel downsampling, and an RGB-vs-grayscale option.
• Fisheye calibration capture (capture_calib.py, calib_config.py) plus per-robot calibration data for both Spots.
• Front-camera color calibration (CCM) — re-fit per-camera color-correction matrices from ColorChecker shots (calibrate_color.py), removing the yellow/pink cast; applied in linear light in the stream.
• Front-view stitching (stitch.py) — depth-based and cylindrical front-camera stitching with exposure/gain seam matching.
• Live RGB comparison viewer (compare_front_rgb.py) — side-by-side and stitched front RGB with live toggles for CCM (c), right→left match (m), and AWB (w).
• Per-frame auto white balance — shades-of-gray AWB applied after the CCM to remove the residual per-camera cast (front-right pink); luminance-preserving, opt-in via --awb.
• Exposure metadata support — reads per-frame exposure/gain from image responses, with tooling to probe it (probe_exposure.py).