Publish corrected Jetson Orin Nano validation evidence

[Kushalk0677]

Summary

• Publish corrected Jetson Orin Nano evidence for CPU ONNXRuntime, PyTorch, and TensorRT routes.
• Apply the corrected detection contract: flat original-image coordinates, class-aware ONNX NMS, RGB ONNX input, and canonical TensorRT COCO class labels.
• Add a TensorRT label-contract test and make the Jetson preflight reject placeholder labels such as class2.
• Rebuild the ready-to-send Jetson validation archive from the public source.

Evidence scope

The TensorRT archive contains the corrected full on-device evaluation, including calibrated runtime profiles, policy-accuracy and VRU-retention stages, manifests, telemetry, and per-rail energy integration. The energy values are rail measurements, not a whole-board energy claim. The included COCO tier-accuracy evidence is contextual; the primary policy evidence remains KITTI replay.

The CPU ONNXRuntime and PyTorch result directories contain their corrected full runtime and retention records. PyTorch also includes a 5,000-image COCO tier-accuracy context run with provenance.

Validation

• python -m unittest -v tests.test_trt_class_labels passes.
• python -m pytest -q tests/test_models.py passes: 68 tests.
• python -m pytest -q tests/test_policy.py passes: 80 tests.
• The rebuilt Jetson archive was extracted and its TensorRT label-contract test passed from the packaged source.

Acknowledgment

Thank you to Evan Leri for preparing the TensorRT label-contract patch, supplying the verification and provenance receipts, and completing the corrected on-device validation evidence preserved in this PR.