[elprofesoriqo]

Summary

• add flowedge-lerobot-package and flowedge_dev pipeline package to build a relocatable LeRobot Diffusion Policy directory
• store the native FlowEdge U-Net plus only the matching RGB encoder and input-normalization state, with source identity and integrity checks
• load local or Hugging Face packages through the normal LeRobot PreTrainedPolicy.from_pretrained path without a dummy root model.safetensors
• document direct lerobot-rollout deployment for SO-101 and add explicit rollout / so101 dependency extras
• build the LeRobot plugin wheel in the release workflow and exercise its package and rollout CLIs in CI

Evidence

• scripts/verify_local.ps1: pass (85 C++ tests, 113 Python tests, conversion/parity gates)
• clean Python 3.12 + LeRobot 0.6.1 plugin suite: 109 passed, 4 skipped, 139 subtests passed
• Sphinx: -W --keep-going pass
• plugin wheel: built, installed, discovered as flowedge, package CLI passed
• real PushT package: 1,050,861,511 bytes; the 44,796,272-byte encoder sidecar avoids 1,006,066,136 bytes of duplicated source U-Net storage
• real package config resolved through LeRobot's plugin registry and loaded the relative native checkpoint without root PyTorch weights

Boundary

This removes deployment packaging and CLI integration friction. It does not claim physical SO-101 task success or Jetson performance; those still require the matched robot run. The generated package defaults to CPU and exact FP32, and approximate CUDA mode remains an explicit post-calibration override.