dbow_kit — WHY (seed)

Seed. Full pitch folds in with a later SKILL.md pass. Evidence: REPORT.md; end-to-end: docs/tutorials/vision_loop_closure.md.

DBoW2 is the canonical bag-of-binary-words place-recognition library behind ORB-SLAM. It has no Python binding and is not on conda-forge — normally that means writing and maintaining a pybind wrapper before you can touch it from Python.

cppyy removes that step: dbow_kit vendors DBoW2's headers, compiles the small .so once (build-dbow2), and mirrors its ORB API (OrbVocabulary, OrbDatabase) directly in Python. Descriptors flow in from cv_kit's C++ cv::ORB as cv::Mats and never leave C++. The result: a real loop-closure detector, assembled and orchestrated from a short Python script, running at C++ speed — the "impossible → possible" case for cppyy.