RoadGlyph + SourceTrace: OCR & RAG on AMD ROCm

Two Dockerized, fully offline solutions for the AMD AI League, both built FROM the required ROCm PyTorch base with Qwen3-VL-8B-Instruct weights baked into the image, so neither needs network access at grading time. RoadGlyph (Mini Challenge 2, OCR) reads road signs and license plates. python3 /app/app.py --input-image <path> writes /app/output/<stem>_output.json with "text" and "confidence". A resident model server keeps the model loaded between images, a watchdog writes a valid fallback JSON before the 30 s limit, and deterministic post-processing applies the transcription conventions for plates, stop signs, speed limits and advisory plaques. SourceTrace (Match 3, RAG) answers questions about a mixed document collection and cites exactly the supporting files. --index builds the index once and --query answers each question. Defensive parsers handle the file types, a vision-language step reads pictures, and retrieval combines BM25 with optional dense search and multi-hop lookup. Answers are grounded verbatim, checked by a separate verification call, and citations are pruned to the exact supporting set. When the documents do not support an answer, it refuses with an empty answer and no citations.