Add locked Recipe onboarding and software acceptance checks

[hootandy321]

Summary

Recipe onboarding previously required manual copies of linked files, environment-dependent next-step commands, and unlocked MicroDuck dependency installation. embodirun example now initializes six linked templates and provides validation, planning, setup, checks, execution, and shutdown through the same implementation as examples/run.sh.

• Preserve the caller's CLI, source root, and selected scene Python in copyable next-step commands, including transient EMBODIRUN_SCENE_PYTHON and paths containing spaces.
• Use dedicated Python 3.12 / uv 0.12.x projects and frozen locks for XLeRobot and MicroDuck, shared by native setup and the production Docker targets. Explicit software profiles support configuration review without robot-owner, model, GPU, or scene dependencies. MicroDuck's default simulation profile retains its full prerequisite checks.
• Add native/Docker software acceptance CI for both Recipes: actual launcher installation, both public entrypoints, printed-command execution outside the checkout, missing-environment diagnostics, non-root file ownership, and package-version parity. XLeRobot also runs both fixture rehearsals and verifies completion metadata and launcher pointer cleanup.
• Keep the existing CPU/lint/docs workflow unchanged. Recipe installer/image checks live in a separate, path-filtered workflow: relevant source, dependencies or container inputs trigger it, and manual dispatch is available. Documentation-only changes do not rebuild Recipe images. The process-cleanup regressions reuse the existing CPU job.
• Preserve XLeRobot hardware-profile setup and Compose's private cross-container shutdown socket. Software cleanup confirmation remains separate from physical stop feedback.
• Finish the command's private process group after graceful leader shutdown, including when the leader has already exited and left descendants. Three real-process regressions verify listener cleanup after completion, parallel failure, and down, while preserving an unrelated process. All three reproduced the leak before the fix.
• Synchronize English/Chinese installation and support documentation. Configuration support for six templates does not imply full-model or physical acceptance of all six.

This branch includes current main, including the RLinf compatibility removal and shared-GPU evaluation changes. PR #7 is stacked on this branch for scenario guides and published software lifecycle evidence.

Type of change

• Bug fix
• New capability (Recipe initialization and setup)
• Documentation
• Refactor with no behaviour change
• Build, packaging, or CI

Verification

For head 4d53ed4515351a45d9edd5f161901381ba72c2f2, the existing CI passed its three jobs (CPU, Ruff, strict bilingual documentation), and the path-filtered Recipe software workflow passed both native + Docker jobs. No additional required status checks were configured; the cleanup regressions run in the existing CPU job.

Check	Observed result
Full CPU suite	Local: 1060 passed, 45 skipped; CI: 1060 passed, 45 skipped
Owned-process cleanup	Three regressions failed before the fix and passed after it; focused Recipe/onboarding suite: 67 passed
Fresh native software setup	Both Recipes installed from their frozen locks with initially empty task download caches and selected Python paths containing spaces
Native/production-container inventories	XLeRobot: 13 matching package versions; MicroDuck: 14 matching package versions; each inventory also checked against its Recipe lock
Container execution	Fresh isolated builder, --no-cache build; checks run as the host UID/GID with --network none
XLeRobot fixture rehearsals	Both CLI and shell runs completed; physical task success remained unverified; temporary launcher JSON pointers removed
Static/documentation checks	Ruff check/format, strict English/Chinese MkDocs builds, and git diff --check passed

Exact CPU/static commands used by CI:

uv sync --frozen
uv run --frozen pytest -q
.venv/bin/ruff check src tests benchmarks examples agents integrations scripts
.venv/bin/ruff format --check src tests benchmarks examples agents integrations scripts
mkdocs build --strict
mkdocs build --strict -f mkdocs.zh.yml

The matrix executes actual setup --mode software, then the selected interpreter runs scripts/check_recipe_software.py --recipe xlerobot|microduck. Its production Dockerfile build and offline container commands are recorded in the linked workflow logs and .github/workflows/recipe-software.yml.

Earlier Thor/Mac and hardware-profile import evidence remains documented in the Recipe guides. Those historical runs are separate from the latest software CI; this update does not claim another hardware-profile or physical-robot acceptance run.

Scope and follow-up acceptance

Ready for review for Recipe onboarding and software configuration/fixture acceptance. No model checkpoint, external scene asset, or physical robot was used in this update; skipped optional tests do not establish those combinations as Tested.

Independent README-based deployment Agent execution, actual student feedback, full simulation/model/CUDA/EGL acceptance, and physical calibration, camera mappings, recorded routes, handover, fresh stop feedback, and delivery acceptance remain follow-up work. Other Recipes' complete dependency locks and installation coverage also remain separate. These results are not reported as completed by software CI.

Process boundaries

• client, deployment, application, devices, model_services remain canonical; services.* compatibility remains available.
• No inference-engine runtime code is imported into src/embodirun; the pinned inference submodule is unchanged.
• Robot/simulator adapters and policy bindings remain separate.
• Robot/model dependencies remain optional and declared in their dependency projects.

Checklist

• Support prerequisites and verification limits are documented in both support matrices; no new verified robot/model combination is claimed.
• English/Chinese onboarding wording and installation/support pages remain parallel.
• No checkpoints, datasets, recordings, credentials, private addresses, or personal data are included.
• Contributions are provided under Apache-2.0 per CONTRIBUTING.md.