技能を、再利用できる学習構造にする。

CUDA runtime and an approved SAM 3.1 checkpoint are required; neither is available in this local build environment.

The official tracker environment/checkpoint has not been installed, and this runtime has no CUDA accelerator.

The official tracker environment/checkpoint has not been installed, and this runtime has no CUDA accelerator.

DINOv3 is conditional on a demonstrated re-identification failure; its approved weights/runtime are not present locally.

Landmarks are analysis previews until a human accepts coverage and identity for each clip; coverage alone cannot validate instructional meaning.

A local pose-only screening found one strong result among four manually bounded bare-hand crops and no strong result among four glove crops; one glove skeleton was visibly misleading. Keep this challenger unavailable for publication until detector, full-clip temporal, source-hash and license gates pass.

Selected 3D-mesh challenger for grip geometry and partial occlusion. It requires an external CUDA run plus approved HaMeR checkpoint and MANO-model terms; no mesh output has been produced in this build.

Technically relevant for temporally consistent 4D interacting-hand motion, but the official code and models are restricted to non-commercial research and also depend on MANO. It is not admitted as a Denso demo dependency.

No repeated domain-labelled detection task has been defined; generic boxes would add visual noise without instructional evidence.

Reviewer- or VLM-proposed keyframes only; masks are not yet propagated through the full clip and remain review-pending.

Reviewer- or VLM-proposed bounded clips only. Start/middle/end identity and drift checks remain review-pending; masks are suppressed outside generated windows.