fix: run local LLM as llama-server sidecar; fix red CI builds

[Aman-Verma-28]

Before Submitting This PR

This is a bug-fix follow-up to #1 (the local-LLM/task feature): it makes the release builds and CI checks pass.

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

Human Written Description

TODO: author to fill in.

Related Issues/Discussions

Follow-up to #1.

What this fixes

After #1 merged, every release build (main-build.yml) failed. Root cause: duplicate ggml/gguf symbols at link time — whisper.cpp (via transcribe-rs) and llama.cpp each statically bundle their own copy of ggml, so linking both into one binary collides on Linux/Windows. (macOS linked by luck; my earlier cargo check never linked, so it slipped through.)

Fix: run the LLM out-of-process. Replaced the in-process llama-cpp-2 crate with a bundled llama-server sidecar:

• On first task generation the app downloads the official prebuilt llama-server for the platform from the pinned llama.cpp release (b9789), extracts it to the app data dir, and runs it on 127.0.0.1.
• LlmManager manages that process (lazy start, idle shutdown) and talks to it over the OpenAI-compatible API, reusing llm_client with the existing response_format: json_schema support.
• llama.cpp is no longer linked into the app binary → the ggml symbol conflict is gone.

Validated locally on macOS: the app links cleanly (was the failure), and the llama-server download + extract + exec works (version: 9789).

Also fixes the other red checks from #1:

• code quality: added the new task i18n keys to all 19 locales (English placeholders; real translations welcome).
• test: the in-memory history test schema now includes the tasks_json column.

Testing

• ✅ cargo build --bin sitrep links cleanly (macOS); cargo test → 69 pass
• ✅ CI test-mock path compiles without transcribe-rs; tsc/eslint clean; check:translations passes 19/19
• ✅ llama-server b9789 download/extract/run verified on macOS arm64
• ⏳ Not verified here: Linux/Windows release builds (should now link since llama.cpp isn't linked in — relying on this PR's CI), the per-platform llama-server asset mapping, and end-to-end task generation against a downloaded model.

Known follow-ups / out of scope

• macOS release signing fails on missing Apple Developer Certificate secrets — pre-existing (every prior main-build failed there), to be fixed in repo settings.
• First-run downloads llama-server silently with no progress UI yet.

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: Diagnosed the CI failures and implemented the sidecar rework. Human review needed for the description, cross-platform binary sourcing, and end-to-end testing.

🤖 Generated with Claude Code

[Aman-Verma-28]

Keeping this open as a reference only — not for merge. Main now builds on Linux + Windows-ARM with the in-process llama-cpp-2 (CPU-only) approach, which is the chosen direction. This PR documents an alternative llama-server sidecar architecture (removes in-process llama.cpp entirely, sidesteps the duplicate ggml/gguf link conflict) in case the in-process build ever causes runtime issues. It conflicts with main's Meeting-mode work and should not be merged as-is.