Do LoRAs and merges trained on non-commercial base weights carry the base license into client work?

FLUX.1 [dev]'s license (read at huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/LICENSE.md) limits the Model and Derivatives to non-commercial purposes yet says Outputs may be used commercially, and also contains a clause about commercial use of data produced by the model that appears to be in tension. Trainers such as ai-toolkit, SimpleTuner and OneTrainer train LoRAs on this base. What is the defensible rule for (a) distributing a LoRA, (b) using it in paid production, (c) using its outputs as training data? Why this is still open: The license text contains internally tensioned clauses and no legal interpretation was sourced; trainer READMEs only say to check each model's terms.