Repair verl OPD semantic contracts - #75

Summary

• make actor/teacher dtype, quantization, attention, and placement choices explicit and compiler-bound
• implement pinned existing student-adapter validation and trainable loading with lineage preservation
• separate logical OPD batch semantics from physical rollout/update token budgets
• centralize placement legality so unknown-size quantized roles require proof instead of selecting swap
• publish machine-generated field-effect evidence for every executable compatibility claim

Why

The v0.9 profile declared several fields compatible while the native runtime either hardcoded a different value or gave the field no documented effect. It could also plan an impossible quantized swap, and lora_adapter_path was not actually loaded.

Impact

The direct-GKD profile now fails closed when placement or adapter provenance is unproven. Existing valid built-in Qwen planning remains executable, and no benchmark result or frozen scientific artifact changed.

Validation

• ruff check and format check
• mypy src/miniverl
• full non-GPU/non-network suite: 2234 passed, 10 skipped, 22 deselected
• coverage: 83.98%
• unit/property suite: 2026 passed
• field-effect artifact: 68/68 eligible fields (100%)
• frozen calculator SHA-256: 53fc1d4d5b7adee09618d77ad62d4086ba56b78569832d6fc7c3bcd5c2695bbc