feat(core): in-process mistral.rs local runtime for qwen3.5-1b

[senamakel]

Summary

• Added a new core-owned local AI runtime in rust-core using mistral.rs with qwen3.5-1b defaults.
• Implemented background bootstrap, model download/checksum handling, warmup, status tracking, and resident in-memory inference.
• Added core RPC methods for local AI status/download/summarization/suggested questions.
• Wired Tauri commands and frontend wrappers to expose local AI debug/interaction controls.
• Switched Tauri thread suggestion generation and context compaction summarization to use the local core runtime (with fallback behavior).

Problem

• Prior flow relied on external local-model orchestration assumptions (e.g., Ollama-style separation), which does not satisfy the requirement that LLM logic run inside rust-core for background execution.
• We needed an in-process local model path that can be initialized once, reused cheaply, and exposed through Tauri only as control/debug endpoints.

Solution

• Introduced local_ai configuration schema and module in rust-core.
• Implemented LocalAiService to:
  • load GGUF via mistralrs::GgufModelBuilder
  • keep Model resident in memory
  • run one warmup request after load
  • provide summarize + suggest APIs
  • download model artifacts when configured
• Added JSON-RPC handlers:
  • openhuman.local_ai_status
  • openhuman.local_ai_download
  • openhuman.local_ai_summarize
  • openhuman.local_ai_suggest_questions
• Triggered background bootstrap at core startup and integrated local summarization into compaction flow.
• Added Tauri command proxies and frontend usage in Settings/Home/thread suggestion paths.

Testing

• yarn -s compile
• cargo check --manifest-path src-tauri/Cargo.toml
• Other checks run (list commands)
  • cargo fmt --all
  • cargo check -p openhuman-core
  • cargo check -p OpenHuman
  • yarn -s tsc --noEmit
  • pre-push hooks (yarn format:check, yarn lint, yarn compile) during push
• Manual validation completed (list scenarios)

Impact

• Desktop/Tauri path now supports an in-process core local model runtime for background summarization/suggestion tasks.
• First-time setup has heavier dependency/build and model artifact requirements (GGUF artifact path/download config).
• No new external model service requirement for this path.

Breaking Changes

• None
• Yes (describe clearly, including migration steps)

Related

• Issue(s): N/A
• Follow-up PR(s)/TODOs:
  • Add explicit model source/bootstrap UX for qwen3.5-1b GGUF download/configuration.
  • Add integration tests for local AI RPC methods and fallback behavior.