Audit the resolved multimodal training contract, not just an example YAML

Audit the resolved multimodal training contract, not just an example YAML. A setting such as LoRA target=all can be filtered by framework freeze defaults, and a model-native chat template can differ from the fine-tuning framework template. Export the exact trained template and preprocessing alongside weights; verify prompt/image masking, answer-plus-EOS supervision and saved/reloaded adapter parity before a long run. When EOS and padding share an ID, mask padding by sequence position/attention, not token equality. These gates establish execution consistency, not task accuracy or equivalence between different architectures.