Gemma 4 Integration Exposes Flutter Limits

This reads less like a model demo and more like a field report on what it actually takes to ship on-device AI: Gemma 4 is useful, but the integration surface is still brittle enough that architecture matters as much as model quality.

• –Direct Kotlin access beat the Flutter wrapper once crashes and even device reboots showed up; abstraction layers are a liability when native inference gets unstable.
• –The single-Engine, one-Conversation-at-a-time constraint is the real concurrency bottleneck, so a global lock is not optional if multiple agents share the model.
• –Multimodal support is real but narrow in practice: JPEG/PNG only, WAV/PCM only, image-size limits, and thinking-mode conflicts all need guardrails.
• –Tool calling and structured output are still soft spots, so production systems need validation, retries, and ground-truth fallbacks for IDs and paths.
• –Thermal throttling turns sustained on-device inference into a systems problem, not just an ML one, which is the main reason this still feels early.