Add bilingual scenario guides and Runtime software lifecycle evidence

[anHappyDog]

Summary

The documentation did not provide clear VLA/VLN/Agent starting paths or distinguish software checks from model and physical acceptance. Parallel English/Chinese scenario guides now connect SO-101, XLeRobot, MicroDuck, and existing Agent workflows to executable Recipes, prerequisites, support boundaries, and a first-use recording procedure.

• Align README entrypoints, quickstart, examples, installation, support matrices, and both navigation trees.
• Add a reproducible SO-101 figure from the existing recorded summary, with a deterministic CI check. Missing raw samples, exact source/checkpoint revisions, and comparable settings remain explicit; the figure is not a new benchmark or a controlled engine comparison.
• Publish three prepared same-Control startup pairs and occupied-port failure/recovery evidence, including sanitized measurements, commands, environment, and lifecycle evidence. One-second probes and assisted preparation support no deployment-performance advantage claim.
• Document independent first-use and six matched deployment-cost comparison procedures, retaining their pending status.

This PR now targets main, after PR #4 merged as da1e02174cc062239cd39ac87e7041ccd0d43cd7. Its head includes that merge commit. Frozen Recipe setup, environment-preserving CLI hints, production Docker profiles, owned-descendant cleanup, and native/Docker software acceptance are already on main. Relative to main, this PR changes documentation, published evidence, and the recorded-figure CI check; core code, tests, locks, Docker setup, and the software acceptance driver are unchanged.

Type of change

• Bug fix
• New robot/model/backend capability
• Documentation
• Refactor with no behaviour change
• Build, packaging, or CI (recorded-figure validation)

Verification

For head 75d594e1560810f6f67b43e5ebf1f6eb75010528, CI passed its three existing jobs: CPU tests, Ruff, and strict bilingual documentation including the recorded figure. The CPU suite passed 1060 tests, with 45 skipped, including three regressions for listener cleanup after command completion, parallel failure, and down.

The separate Recipe software workflow is path-filtered and did not run for this documentation/evidence diff. The only additional step relative to PR #4 is the inexpensive recorded-figure check inside the existing docs job. Installer/image acceptance remains available for relevant source/dependency/container changes or manual dispatch; PR #4's matching implementation passed both jobs in its Recipe software run.

Local Ruff check/format, strict English/Chinese builds, figure validation, and git diff --check passed. PR #4's matching core/tests also passed the full local CPU suite (1060 passed / 45 skipped).

.venv/bin/ruff check src tests benchmarks examples agents integrations scripts
.venv/bin/ruff format --check src tests benchmarks examples agents integrations scripts
.venv/bin/python benchmarks/engine-comparison/generate_plot.py --check
.venv/bin/mkdocs build --strict
.venv/bin/mkdocs build --strict -f mkdocs.zh.yml
git diff --check

PR #4's Recipe jobs use actual frozen native launcher installation and production Dockerfile builds with a fresh isolated builder and --no-cache. Both public entrypoints run as the caller UID/GID; container checks use --network none. XLeRobot and MicroDuck inventories are verified against their locks and compared between native/container environments. XLeRobot also executes both software fixture rehearsals while retaining unverified physical results.

The prepared Thor report records its own measured source revision, assistance, workload, and uncertainty. Earlier MicroDuck Thor installation evidence remains scoped as assisted native/cached Docker work, with initial failures and incomplete full-profile installation preserved; latest software CI does not replace those historical facts.

Scope and follow-up acceptance

Ready for review for bilingual scenario documentation and published prepared software lifecycle evidence. No newly verified model/robot combination or measured Runtime deployment-cost improvement is claimed.

Full simulation/import/CUDA/EGL and model/scene/navigation acceptance, a native LeRobot protocol adapter and six matched deployment-cost comparisons, independent human/classroom first use, and physical-robot acceptance remain follow-up work. Software CI and prepared Control trials do not complete them.

The PR #4 dependency has landed, and this branch is synchronized with its merge on main. Required human review and branch protection remain in effect.

Process boundaries and checklist

• Canonical Runtime domains and compatibility imports remain intact; model runtime code remains outside src/embodirun.
• Robot/model dependencies remain optional; this review diff leaves the root lock and pinned inference submodule unchanged.
• English/Chinese guides agree on prerequisites, evidence, and unverified combinations; both strict navigation builds pass.
• Support matrices retain actual verification boundaries rather than inferring Tested status from code presence.
• No checkpoints, datasets, recordings, credentials, private host addresses, or personal data are included.
• Contributions are provided under Apache-2.0 as described in CONTRIBUTING.md.