eval: FROZEN_LABEL env var to avoid overwriting existing baselines

eval: EVAL_TEMP env var to measure latent policy preference

eval: auto-detect LoRA base model from adapter_config.json

in-distribution eval support: EVAL_PHASE env var

phase-2: SKIP_FROZEN + TRAINED_LABEL for re-eval against new checkpoint

phase-g: T4 eval wrapper for Frozen + Trained Qwen on eval pool