Runtime conditioning scale reduction must not be carried into persisted model or LoRA weights. Keep named merger presets at 1.0 and reserve scaling for explicit custom configuration.
- build generic and LoRA worklists from the union of inputs
- copy single-source layers exactly and merge available providers
- keep logical LoRA groups canonical, atomic, and collision-safe
- retain mismatch, dtype, quantization, streaming, and OOM contracts
- preserve direct diffs, precision, and guarded low-bit tensors
- stream paired CWB layers with atomic CPU OOM fallback
- save generated merge artifacts under models_dir
silveroxides