Release v0.2.6 lifecycle and privacy correctness - #25

Summary

• make training, evaluation, checkpoint save/load, and destructive close share one non-blocking operation owner
• restore the exact previous model train/eval mode after successful or failed evaluation
• harden portable semantic-key, credentialed-URL, and embedded absolute-path redaction without hiding known benign metadata
• finalize the focused v0.2.6 release metadata; no algorithm, environment, benchmark, model family, adapter, or frozen result changed

Validation

• 209 focused lifecycle/privacy regressions
• 1,348 non-GPU/non-network tests, 87.46% branch coverage
• 5 GPU tests on RTX 4080; 3 network tests
• Python 3.10–3.13 core: 1,058 tests each
• minimum and latest training boundaries: 2 smoke + 133 compatibility tests each
• WSL2/Linux concurrency slice: 76 tests
• Ruff, format, mypy, actionlint, Markdown links, generated-file comparisons
• wheel/sdist + Twine; extracted-sdist 1,348-test suite; matching 78-file wheel inventory
• clean core and [train] wheel installs with demo/inspect/report/eval

Frozen benchmark SHA-256 remains 53fc1d4d5b7adee09618d77ad62d4086ba56b78569832d6fc7c3bcd5c2695bbc.