Harden OPD freshness, benchmark provenance, and protocol teachers - #4

Summary

• make strict OPD freshness the default and label multi-update rollout reuse as online distillation with replay
• replace ambiguous distillation CE naming with sampled-token NLL semantics
• add benchmark schema v2 with pre-allocation structured-diff validation, portable config/checkpoint provenance, cumulative accounting, and v1 read compatibility
• add standard frozen PEFT teacher adapter validation/export, policy-competence gates, and a prespecified Qwen3-1.7B protocol-teacher recipe
• tighten full-vocabulary and temperature-scaling claims, add numerical sweeps, and update CI/dependency/release controls for v0.2.0
• publish the complete five-arm, two-seed RTX 4080 result as schema-validated JSON, Markdown, and a source-bound SVG

Measured result

Candidate A reached 100% strict held-out tool-policy success on 24 tasks and passed the prespecified 50% gate, so candidates B and C were not run. In the equal-optimizer-update benchmark, both seeds produced the same strict-success ordering:

Protocol alignment prevents the observed collapse and ties SFT; it does not establish an OPD-over-SFT advantage. The legacy adverse result and transcript diagnosis remain published.

Published result SHA-256: 53fc1d4d5b7adee09618d77ad62d4086ba56b78569832d6fc7c3bcd5c2695bbc.

Validation

• ruff check . and ruff format --check . — 148 files
• mypy src/miniverl — 73 source files
• full offline/CPU suite — 966 passed, 4 deselected, 86.55% branch coverage
• pytest -q -m gpu — 4 passed on RTX 4080
• Transformers 4.51.3 offline Qwen3/PEFT/config boundary — 117 passed
• python -m build and python -m twine check — 0.2.0 wheel and sdist passed
• fresh wheel-only environment — help/version/doctor passed with torch absent
• fresh [train] environment — miniverl demo --fast completed end to end
• 41 Markdown files — 117 local links and anchors valid

No tag, GitHub release, or PyPI publication is part of this PR.