Harden trainer lifecycle and v0.2 release provenance - #5

Summary

• make OPDTrainer.close() destructive and idempotent, including construction-failure cleanup, post-close guards, optimizer/target/backend teardown, and CUDA allocator release
• isolate every benchmark cold start and arm inside its own trainer lifetime
• separate declared scientific diffs, runtime resolution, and harness bookkeeping in additive schema-v2 provenance fields
• publish and pin the verified protocol-teacher adapter, while retaining an explicit local/offline config
• correct legacy reproduction instructions and document the historical reserved-memory caveat without modifying measured results

Validation

• ruff check .; ruff format --check .; mypy src/miniverl (73 source files)
• CPU/network-excluded suite: 981 passed, 6 deselected
• GPU suite: 5 passed, 982 deselected
• sequential CUDA lifecycle regression: both 32,330,752-byte live probes returned allocated/reserved memory to 0/0 at close() return
• Transformers 4.51.3 and 5.14.1: the same 118-test offline Qwen3/config slice passed on each
• final wheel and sdist built; both passed twine check
• fresh wheel-only core install had no torch; fresh [train] wheel install completed demo, inspect, and report
• fresh remote branch clone completed the documented full source-install demo in 43.5 seconds / 132 optimizer steps
• 39 Markdown files: 117 local links/anchors and 11 external links passed
• protocol adapter verified from Hugging Face Hub at revision 23323751318135484c06c043b1f9b9e7016dd89f

The published benchmark result remains byte-identical at SHA-256 53fc1d4d5b7adee09618d77ad62d4086ba56b78569832d6fc7c3bcd5c2695bbc.

Release status

No tag or GitHub Release is created. PyPI publication remains externally blocked: miniverl has no PyPI project/trusted publisher yet, and this GitHub repository has no pypi environment. The release workflow therefore remains hard-disabled with if: false.