act_flip_v1_counterroll

v1 expert with the wrist pre-rolled so the flip counter-rotates against the base swing

Jacob's from-scratch ACT (scripts/act.py, scripts/train_act.py in https://github.com/JaHeRoth/robot-learning). Checkpoints uploaded: best = step 7500, final = step 30000. Raw torch.save dicts with model_state, opt_state, stats.

step	success	95% CI
2500	0.384	[0.342, 0.427]
5000	0.782	[0.744, 0.816]
7500	0.948	[0.925, 0.964]
10000	0.848	[0.814, 0.877]
20000	0.810	[0.773, 0.842]
30000	0.858	[0.825, 0.886]

500 held-out seeds (800000-800499), n_action_steps 40, horizon 450