Release v0.2.5 correctness hardening - #22

Summary

• make Calculator, JSON-navigation, and SQLite verification total over arbitrary model strings; classify non-finite SQLite answers as malformed
• redact semantic secret keys, URL userinfo, Windows/UNC/private POSIX paths across portable artifacts
• publish precise ready/running/terminal manifests, keep automatic reports and standalone checkpoint selection under one run lock, and reject public eval/checkpoint calls during training
• use verifier-format-valid protocol-v2 final examples while preserving protocol-v1 bytes
• rewrite nested linked images and reject relative project targets in PyPI Markdown and wheel metadata
• finalize the focused v0.2.5 release metadata without changing algorithms, environments, models, adapters, or benchmark data

Regression evidence

The pre-fix tests reproduced 23 verifier/protocol failures, 16 privacy failures, 5 lifecycle/API failures, the automatic-report lock race, two standalone-eval lock races, and the nested License target. The focused post-fix surface passes 798 tests.

Validation

• ruff check ., ruff format --check ., mypy src/miniverl, actionlint
• 1278 non-GPU/non-network tests; 87.37% branch coverage
• 5 local RTX 4080 GPU tests; 3 network tests
• Python 3.10-3.13 torch-free core: 1021 tests on each version
• minimum and latest training bundles: 2 no-network training smokes each
• Transformers 4.51.3 and 5.14.1: 133 compatibility tests each
• final wheel/sdist build and Twine check; extracted-sdist suite: 1278 tests
• clean core install plus clean [train] demo/inspect/report/eval
• wheel and sdist-built wheel inventories match (78 runtime files); zero relative metadata targets
• frozen benchmark SHA-256 remains 53fc1d4d5b7adee09618d77ad62d4086ba56b78569832d6fc7c3bcd5c2695bbc

Windows spawn-based concurrency tests pass locally. WSL was unavailable because its VHDX returned ERROR_SHARING_VIOLATION; the PR Ubuntu jobs are the Linux multiprocessing evidence gate.