Capture the design for a future `ANALYSIS_BACKEND=hybrid`: ocrs extracts
text, the vision LLM does grounding (tags/scene/safety), with the two kept
memory-exclusive on the Pi by reusing the vision-manager's existing
crawl-mutex pattern (OCR is a third RAM-competing, higher-priority
workload). Design only — not implemented.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
