feat: local on-device LLMs + task extraction from transcripts

[Aman-Verma-28]

Before Submitting This PR

⚠️ Heads up (feature freeze): This is a new feature, not a bug fix. Per the template and CONTRIBUTING, new features need community support gathered in Discussions before merge. No discussion has been opened yet — please start one and link it under Community Feedback, or treat this as a proposal/draft.

Please confirm you have done the following:

• I have searched existing issues and pull requests (including closed ones) to ensure this isn't a duplicate
• I have read CONTRIBUTING.md

If this is a feature or change that was previously closed/rejected:

• I have explained in the description below why this should be reconsidered
• I have gathered community feedback (link to discussion below)

Human Written Description

TODO: author to fill in.

Related Issues/Discussions

Fixes #
Discussion:

Community Feedback

Not yet gathered.

What this PR does

Adds four open-weight, on-device LLMs — DeepSeek (R1-Distill), Qwen 2.5, Google Gemma 2, and Mistral — as downloadable GGUF models that run fully offline via llama.cpp, alongside the existing voice models. On top of that, a new "Tasks" feature extracts a structured task list (title / description / priority) from any transcript in the History view, with no API keys and no network at inference time.

Backend

• New ModelKind { Transcription, Llm } so LLM models share the download/verify/storage machinery but never appear in the voice model selector.
• LlmManager mirrors TranscriptionManager (lazy load, idle-unload, panic-safe). Uses the GGUF's embedded chat template, strips DeepSeek-R1 <think> blocks, and constrains output to valid JSON with a GBNF grammar derived from the task schema.
• Models download directly from Hugging Face with pinned SHA256 (small + standard size per family).
• tasks_* settings, task-prompt CRUD commands, and a tasks_json history-DB column (append-only migration) so generated tasks persist.

Frontend

• "Tasks" settings panel (model picker reusing ModelCard + prompt editor), an enable toggle, a useLlmModels hook, and a TaskListPanel with generate/regenerate, priority badges, and copy.

Testing

• ✅ cargo check clean (0 errors/warnings from new code); cargo test tasks:: → 4/4 pass (JSON/<think>/bare-array/empty parsing)
• ✅ tsc --noEmit clean; eslint src/... clean; Prettier applied
• ⚠️ Build note: native deps (whisper.cpp + llama.cpp) must compile with Apple Clang, not Homebrew GCC. Verified locally with CXX=/usr/bin/clang++ CC=/usr/bin/clang (and without CMAKE_POLICY_VERSION_MINIMUM). CI/tauri-action uses the standard toolchain, so this should be fine there.
• ⏳ Not yet done: end-to-end manual run (download a model → generate tasks → restart to confirm persistence). Help with additional testing welcome.

Translations

Only en/translation.json has the ~37 new keys; the other 19 locales fall back to English at runtime. check:translations will report them missing until translators add them (per CONTRIBUTING_TRANSLATIONS.md).

Screenshots/Videos (if applicable)

AI Assistance

• No AI was used in this PR
• AI was used (please describe below)

If AI was used:

• Tools used: Claude Code (Opus 4.8)
• How extensively: Substantial. AI explored the codebase, designed the approach, and wrote the implementation (Rust backend, React/TS frontend, bindings, i18n) and tests. A human should review and complete the human-written description, gather community feedback, and run end-to-end manual testing before merge.

🤖 Generated with Claude Code