**Scaled training plans** add default-off CUDA/TRL profiles for 7B QLoRA RLVR and sharded 32B/72B distillation across Python and TypeScript.
**Training scale metadata** records device count, sharding, memory budgets, quantization, parameter count, and deployment VRAM for registry gating.
**TypeScript CLI parity** makes `--scale-profile` preserve profile backend/base/mode and documents every scale-related train flag.
