The previous release served `absolute_eef10` (absolute EEF pose, with
osc_position_scale=0.05 / osc_rotation_scale=0.5 in the stats), while the
LIBERO evaluation client in benchmarks/libero consumes the native normalized
delta command `[delta_xyz3, rot6d(Exp(delta_rotvec3)), gripper_open1]` without
composing with the current pose and without applying the controller scales.
This checkpoint is trained on the delta representation, matching both that
client contract and assets/benchmark_data/libero (conversion.json declares
output_representation=native_delta_eef10). normalization_stats.npy is now
byte-identical to the canonical benchmark_data stats file (keys: eef,
eef_state).
Measured on LIBERO with the official protocol (50 trials/task, seed 42),
sampled one task per suite: spatial 50/50, goal 49/50, object 50/50,
long 48/50 -> 197/200 = 98.5%.
