"""Training: pre-encoding rendered windows into model inputs, and causal LoRA training (`cwm` environment)."""