Publish TrainerRank checkpoint slots

[bradhilton]

Summary

• retain validated adapter metadata on trainable TrainerRank checkpoint slots
• export a selected dynamic slot through the existing distributed vLLM LoRA codecs without mutating base LoRA tensors
• preserve legacy alpha-only loading and static publisher behavior
• cover dense exact export and static/dynamic Qwen 3.5 packed-MoE codec paths

Why

Caladan coordinates cache upload, refresh, and W&B commits, but ART owns the live trained tensors and model-specific export codecs. Clean ART main had no supported way to serialize a dynamic checkpoint slot, so Caladan could only fail closed rather than risk publishing stale base A/B tensors.

Validation

• TrainerRank validation suite: 43 passed before the final legacy-compatibility assertion was added
• focused Ruff: passed on the final diff
• focused ty: exit 0 on the final diff; existing torch.distributed attribute warnings remain
• dense and packed-MoE codec tests are included, but local collection requires Megatron Bridge/Transformer Engine and a CUDA-capable environment; the CPU workstation cannot build DeepEP or import Transformer Engine