Deep dive into Google's newly expanded Gemini Nano on-device capabilities announced at I/O 2026, specifically the function calling and structured JSON output features. Cover the practical architecture for building offline-capable AI agents on Android: how to define tool schemas within the on-device token budget, handle the constrained context window (32K vs cloud models), implement reliable JSON extraction from a quantized model that hallucinates more than its cloud counterpart, and the WorkManager + Room pipeline for queuing actions when connectivity returns. Include real benchmarks comparing Gemini Nano function calling accuracy vs cloud Gemini Flash for common mobile tasks (calendar events, contact lookup, message drafting).