v0.6.1 visual integrity and bridge correctness - #40

What this changes

This focused v0.6.1 release replaces four misleading Alignment Lab scatterplots with three data-bound row/forest views, rebuilds the bridge diagram as verified layers plus a prominent untested layer, adds real-browser documentation gates, modernizes and versions the docs site, and makes import/export readiness semantics fail closed.

The README is reduced from 32.9 KB to 8.1 KB and the Chinese README from 20.0 KB to 6.0 KB while keeping the negative calculator, RecoveryBench and Alignment Lab conclusions linked and explicit.

Bridge contract

• import-verl classifies every source field as exact, derived, informational_only, requires_user_confirmation or unsupported.
• Incomplete imports emit import-report.json plus non-executable imported.template.yaml with status needs_user_input.
• No calculator fallback, unqualified same-base teacher, NaN/infinity or unresolved interpolation is accepted.
• Runnable recipes require explicit environment, teacher, loss and schedule choices and pass RunConfig validation before publication.
• export-verl writes launch.template.sh and reports artifact, upstream smoke, reward, launchability, distributed-execution and semantic-parity states separately.
• The current reward scaffold remains fail closed and launchable: false.

Visual and docs evidence

• New Alignment Lab figures expose all three seeds, exact axes, zero baseline, N/A query ratios and external-benchmark coverage.
• Desktop and mobile bridge diagrams separate local runtime, portable bundle, pinned v0.8.0 / 7aed6b23 smoke and Distributed execution: NOT TESTED.
• Material 9.7.7 provides stable/dev selection, search, dark/light modes, responsive tables and navigation.
• Playwright inspects 1440x900, 1024x768, 820x1000 and 390x844 across the five required pages; final Linux run 30856925490 passed and uploaded 20 manually inspected screenshots.
• Future external alignment endpoints are tracked separately in Roadmap: preregister real external alignment benchmarks #39.

Validation

• ruff check and format: passed
• mypy: 103 source files passed
• actionlint 1.7.12: passed
• non-GPU/non-network: 1563 passed, 6 deselected, 86% branch coverage
• GPU: 5 passed
• network: 3 passed
• extracted sdist: passed in final build run 30856925481
• clean core and [train] wheel installs plus a real toy demo: passed
• pinned verl source smoke: config merge, PEFT/safetensors/Parquet, privacy and hashes passed; launchable: false; distributed execution not tested
• strict MkDocs, Markdown/link, generated-byte and Playwright visual gates: passed
• exact v0.6.1 build and Twine check: passed

Immutable evidence

No benchmark JSON/JSONL or committed bridge smoke record changed. All 12 audited frozen hashes were rechecked. The calculator result remains SHA-256 53fc1d4d5b7adee09618d77ad62d4086ba56b78569832d6fc7c3bcd5c2695bbc; Alignment Lab result/task hashes remain 584752dccb91654109c357b8ebb12681a12a9c1476a9ba539dd35e4d860a22ef and 8d7fc723436d7377d196fc44046d960e3cb7f0aa81e03d49ef05b627eb84630f.

Checklist

• User-visible docs updated in English and Chinese.
• New behavior has failing-first regression coverage.
• No new scientific experiment or distributed execution claim.
• No base-install dependency added.
• Privacy and machine-path scans passed.
• Exact 0.6.1 release metadata finalized.
• Final CI, build, docs and pinned-profile jobs green.