Codex/flutter gemma integration points

[dev07060]

Summary

• Add MobileRagVectorStore as a small adapter surface for external embedding / LLM runtimes such as flutter_gemma.
• Export the new vector-store facade from the package public API.
• Add native tests for precomputed embedding storage/search and SourceRagService.dispose() lifecycle cleanup.
• Harden log-stream cleanup with a timeout so dispose does not hang.
• Keep flutter_rust_bridge pinned to 2.11.1 for the current generated binding/runtime surface.

Improved

This creates a concrete coupling point for flutter_gemma-style integration without claiming that mobile_rag_engine bundles or replaces an LLM runtime. The useful bridge is now clearer: external runtime produces or consumes embeddings/messages, while mobile_rag_engine owns local storage/search behavior.

Trade-off

MobileRagVectorStore is intentionally narrow. It maps one external document to one source/chunk row, so it is useful as an adapter/proof surface, but it does not yet expose the full parsing, chunking, hybrid retrieval, source hydration, and context assembly strength of the main RAG pipeline.

Verification

• flutter analyze
• flutter test

Result: all checks passed, including 80 tests.

Follow-up

• Build a flutter_gemma companion/demo layer around the stronger context path:
  searchHybridWithContext(..., hydrationMode: contextOnly) -> prompt/message bridge -> flutter_gemma chat generation.
• Add a small example app or doc showing the exact handoff boundary.
• Avoid broad “deep integration” language in promotion; position this as a practical local-RAG retrieval bridge for on-device LLM apps.