Release v0.7.0 early-stop evidence - #53

Summary

• publish the schema-valid v0.7 external-alignment checkpoint-selection early-stop result
• preserve original selection evidence, add a metadata-only fallback-lineage correction, disclose task-identical selection suites, and publish 512 privacy-safe task rows
• add responsive data-bound figures, the scoped study page, bilingual entry points, and miniverl pilot --study-result
• harden lineage capture and make evidence regeneration byte-identical from both the checkout and extracted sdist
• finalize v0.7.0 package and release metadata

Scientific boundary

Both declared lineages and all eight candidates measured 0/64 retained JSONNav utility against the unchanged 20–90% gate. No checkpoint was selected. Teacher/judge qualification, SFT/DPO/KD/OPD continuation, method comparison, and the reserved final test were not run. The final test remains not accessed with zero tasks scored. Granite Guardian values remain unqualified diagnostics and no positive OPD result is claimed.

No pre-existing frozen result JSON/JSONL changed. The calculator artifact remains SHA-256 53fc1d4d5b7adee09618d77ad62d4086ba56b78569832d6fc7c3bcd5c2695bbc.

Validation

• Ruff check and format: pass
• mypy over src/miniverl and scripts: 146 files pass
• actionlint: pass
• CPU/non-GPU/non-network: 2,098 passed, 6 skipped, 21 deselected; 86.35% combined branch coverage
• available RTX 4080 suite: 8 passed
• network suite: 15 passed
• focused evidence/package contracts: pass, including byte-identical regeneration
• extracted sdist: 2,096 passed, 8 skipped, 21 deselected
• mkdocs build --strict: pass
• browser visual gate: 36 rendered SVG instances across 1440×900, 1024×768, 820×1000, and 390×844; screenshots manually inspected
• package build and Twine: pass
• Markdown links, text integrity, release-state, attribution, privacy, and git diff --check: pass