Release v0.7.1 product correction - #56

What changed

• reorder the English, Chinese, PyPI and docs landing pages around the installable single-GPU runtime, bounded verl bridge and measured systems evidence
• move preserved research outcomes below the primary user journey without changing their negative conclusions
• add wheel-packaged alignment-external-v1 result, schema, preregistration and 512 task rows
• add miniverl evidence show/validate and pilot --builtin-study so pip users do not need a checkout
• finalize v0.7.1 release metadata while leaving v0.7.0 and all frozen evidence immutable

Validation

• Ruff check/format: 332 files
• mypy: 147 source files
• actionlint 1.7.12
• CPU/non-GPU/non-network: 2110 passed, 6 skipped, 21 deselected; 86.19% branch coverage
• RTX 4080 GPU: 8 passed
• network: 15 passed
• extracted sdist: 2107 passed, 9 skipped, 21 deselected
• strict MkDocs and browser checks: 36 SVG instances across 1440, 1024, 820 and 390 px viewports
• clean core and [train] installs; core keeps torch absent and validates all packaged evidence
• wheel/sdist build and Twine check
• no benchmark, evidence or preregistration bytes differ from v0.7.0
• owner-only source-author and forbidden-trailer audit

Boundary

This maintenance release corrects product hierarchy and installed evidence UX. It does not implement or claim the planned v0.8 verl-shaped OPD runtime, distributed execution or a new scientific result.