[ZJLi2013]

Problem

New users had no single place to learn how RoboSmith works or to get from a fresh GPU node to a first run.
docker/README.md built docker/Dockerfile, which no longer exists, and the steps for GraspGen weights,
the gfx950 GraspGen layer, and the rocRobo sidecar were not written down anywhere in the repository.

Change

• docs/design.md: data flow, layer contracts, and the key concepts (asset affordances and metadata, CAP
  scenarios, the grasp filter pipeline, segment-level motion routing, the collision world, articulation, outputs),
  plus current limits.
• docs/quick-start.md: host layout → image per GPU arch → GraspGen weights → rocRobo sidecar →
  run_scenario.sh → outputs, and troubleshooting for cold compile, dead serve, and straight-line fallback.
• docker/README.md and the Dockerfile.gfx950 header build from the files that exist.
• Both READMEs link the two docs from their quick-start section; CHANGELOG entries.

Layer: docs and runtime images. No code changes.

Evidence

• Every repository path, identifier, and relative link cited in the new docs was checked against this branch
  (git ls-files / git grep -w); none missing.
• Defaults quoted in the docs were read from the code: rocrobo_dev, /rocrobo/.jax_cache,
  ROCROBO_REQUEST_TIMEOUT_S=300, the *_rocm.pth preference in graspgen_wrapper.py, the dataset path in
  run_generated_scenario.py, and the launcher defaults in run_scenario.sh.
• The quick-start commands were not re-run end to end on a fresh node for this PR.

AI use

• Tool/model: Cursor agent (Claude)
• Extent: documentation (drafted both docs from the code and prior notes), review (path / identifier audit)
• Human verification: maintainer review of the rendered docs before merge