Drop barrel, add held-out stop-sign distance, add the GPU runbook

Make setup.sh say what it is doing

Add a Pi 5 runbook, and correct the NCNN diagnosis

Add a portable evaluation package, and drop the transcription files

Fix two cross-dataset convention bugs found by scoring against ground truth

Add COCO and AV2 fetchers that pull only what the pipeline reads

Phase 1: unified data pipeline, QA and training