sancov/libero_recap_snapflow-test — SmolVLA+RECAP on LIBERO (10-step weights run with snapflow_enabled)

Reference checkpoint for TensorRT conversion accuracy benchmarking in ros-fhnw-autonomy (docker/benchmark/). It is published so the on-device benchmark job can load real trained weights: half precision saturates on the outlier activation channels that only appear in trained weights, so an FP16 engine checked against a randomly initialised policy passes a check it would fail in deployment.

• Policy: smolvla_recap (10-step weights run with snapflow_enabled)
• Validated against: sancov/smolvla_recap_libero_spatial
• Inputs: observation.images.image, observation.images.image2 (3x256x256), observation.state (8)
• Output: action (7), chunk size 20

These are the same tensors as sancov/libero_recap_final-test. This variant only sets snapflow_enabled: true in config.json; it is not a SnapFlow-distilled model. For that, use sancov/snapflow_ctpf08-test. Kept because earlier benchmark runs used this configuration.

Used as the eager reference in the eager-vs-TensorRT comparison; not intended as a general-purpose LIBERO policy release.