Add measured sampled-k1 OPD profile - #80

Summary

• add the closed �erl-opd-v0.8-single-gpu-pg-k1-v1 profile with pinned k1 estimator and vanilla policy-loss semantics
• bind rollout actor log-probabilities, sampled teacher targets, policy versions, cache/checkpoint identity, and exact verl v0.8 scalar/gradient conformance
• support profile-aware export, doctor, and transactional materialization without top-k requirements
• publish RTX 4080 systems evidence with 32 distinct prompts, 8 strict updates, and exact update-4 resume
• keep the published direct-GKD profile schema and generated compatibility bytes unchanged

Validation

• 2286 passed, 7 skipped, 22 deselected; 83.66% coverage
• GPU: 8 passed
• network: 16 passed
• pinned verl PG conformance: scalar, metrics, gradient, optimizer step
• RTX 4080 PG workload: 3.1914 GiB peak reserved; exact interruption/resume
• materialized PG scale-out: pinned config merge and model/data load smoke passed; distributed execution not tested
• ruff, format, mypy, actionlint, strict MkDocs, Playwright 4-viewports, build and twine passed
• frozen calculator SHA-256 remains 53fc1d4d5b7adee09618d77ad62d4086ba56b78569832d6fc7c3bcd5c2695bbc

Scope

This is systems and semantic-conformance work. It does not evaluate task quality, compare algorithms, implement PPO/GRPO, or claim distributed verl execution or full compatibility.