Measure the full SmolLM2 OPD recipe - #81

Summary

• add a pinned SmolLM2-360M/1.7B direct-GKD recipe and reproducible eight-update workload driver
• publish RTX 4080 evidence for 32 distinct prompts, exact update-4 resume, PEFT reload, and materialized scale-out validation
• record a separate Ubuntu 26.04 WSL2 plan/probe/one-update/PEFT smoke on the same physical RTX 4080
• filter model materialization downloads to the top-level Transformers snapshot payload so auxiliary ONNX exports cannot poison or inflate a bundle

Measured scope

• Windows RTX 4080: 8 updates, 1.4961 GiB peak reserved, 23.1082 s first update, exact resume
• WSL2 RTX 4080: measured probe plus one update, 1.4043 GiB peak reserved, 9.5106 s first update
• exact snapshot materialization: config parse, Parquet, PEFT, tokenizer, CPU model load and tiny forwards passed
• distributed execution and task/alignment quality remain untested

Validation

• focused workload, materialization, conformance, packaging and frozen-hash tests
• ruff check and format check
• mypy src/miniverl scripts
• text integrity
• mkdocs build --strict
• full repository CI pending on this PR