Publish the realistic RTX 4080 OPD workload - #72

What changed

• adds a reproducible 64-prompt, 8-update Qwen3 OPD developer workload on one RTX 4080
• fixes Parquet resume so the saved cursor reconstructs the exact epoch and row offset
• verifies byte-identical uninterrupted/resumed trajectories, adapter tensors and optimizer tensors
• publishes data-bound desktop/mobile systems figures and concise English/Chinese documentation
• qualifies pinned Apache-2.0 SmolLM2-360M/1.7B with a one-update compatibility smoke

Measured scope

The Qwen3 run consumed 32 distinct prompts, used a 64-token response bound and reached 3.1914 GiB peak reserved VRAM. Median rollout, teacher-scoring and update times were 9.7200, 0.4864 and 2.3260 seconds. The SmolLM2 check is explicitly compatibility-only, not a second full recipe or quality result. No distributed execution or OPD superiority is claimed.

Validation

• 2222 non-GPU/non-network tests passed; 83.95% coverage
• 8 GPU tests passed; 15 network tests passed
• pinned verl v0.8.0 conformance passed (2 numerical + 1 materialization)
• Ruff, format, mypy, actionlint, text integrity, release state and Markdown links passed
• strict MkDocs and 44 browser-rendered SVG instances across 1440/1024/820/390 px passed
• wheel/sdist build and Twine passed; clean core/train installs and extracted-sdist tests passed
• frozen calculator SHA-256 remains 53fc1d4d5b7adee09618d77ad62d4086ba56b78569832d6fc7c3bcd5c2695bbc