Add the single-GPU verl-style OPD plan and run path - #61

Summary

• add weight-free miniverl plan and validated miniverl run workflows for the pinned verl v0.8-style single-GPU OPD profile
• package a pinned Qwen3 0.6B/1.7B NF4 quickstart, portable Parquet sample command, independent response bounds, resume support, and standard PEFT export
• publish a scoped RTX 4080 runtime-conformance record and quickstart without making alignment-quality or distributed-execution claims
• retain the established environment-backed cycle metric contract while adding prompt-runtime phase timings

Why

The compiler and loss conformance work had no cohesive user path from source Parquet to a resumable local run and loadable adapter. This change exposes that path while keeping every scientific result frozen and making measured versus estimated planning fields explicit.

Validation

• ruff check .
• ruff format --check .
• mypy src/miniverl
• actionlint
• pytest -q -m "not gpu and not network" --cov=miniverl — 2174 passed, 8 skipped, 85% branch coverage
• pytest -q -m gpu — 8 passed
• pytest -q -m network — 15 passed
• official verl v0.8.0 loss/metric/gradient conformance — 2 passed at 7aed6b230776f963fa09509c10d9c3a767d1102c
• mkdocs build --strict
• package build and twine check
• frozen calculator result SHA-256 remains 53fc1d4d5b7adee09618d77ad62d4086ba56b78569832d6fc7c3bcd5c2695bbc

Scope boundary

The measured record proves one local single-GPU runtime and artifact path. It does not evaluate alignment quality, compare OPD with SFT/DPO/KD, or claim distributed verl execution.