Stage 1: Florence-2-base + text-encoder LoRA (rank 8 alpha 4 dropout 0.2) trained
to predict the target bowl's pixel coordinates given (image, prompt). Within-condition
val split. slot_acc ~0.96.
Stage 2: load Stage 1 ckpt, attach DiT action head, joint train with
total = action_loss + 0.1 * pixel_loss (Stage 1 head kept as regularizer).
Florence base frozen, LoRA adapters fine-tuned, DiT trained from scratch.
Package: text LoRA merged into Florence base weights so this checkpoint
loads into vanilla FlowerVLAPolicy without LoRA support.
