feat: bundled engine runner and model library

[quiet-node]

Overview

Phase 2 of the local inference engine, part 1 of 3 (stacked: this PR, then #218, then #219). Thuki gains a bundled llama.cpp llama-server sidecar and a local model library, so the app can serve chat without any external runtime. This PR lands the engine process management and the model storage/download machinery; nothing routes chat through it yet (that is the next PR in the stack), so user-visible behavior is unchanged.

What changed

• Engine runner (src-tauri/src/engine/): a pure residency state machine (Stopped/Starting/Loaded/Stopping/Failed) driven by an actor over a bounded command channel. At most one llama-server process ever exists; a model or context-size change is a new Target and always kill-then-spawn; the process is killed and its exit awaited on app quit. Health is polled over loopback; an optional idle_unload_minutes config field (new, default off) unloads after inactivity.
• Model library (src-tauri/src/models/ split into submodules): SQLite manifest of installed models, content-addressed blob store (blobs/<sha256>), resumable downloader (Range requests against pinned Hugging Face revisions, sha256 verification, disk-full/offline/checksum error classification, single download slot), and a curated starters registry (Gemma 3 4B/12B, Phi-4 14B; pinned repo revisions and digests; RAM-fit tiers).
• Nine Tauri commands for the model library (list/download/cancel/resume/discard/delete and friends) emitting a typed DownloadEvent channel contract.
• Build wiring: scripts/ensure-llama-server.ts fetches the pinned llama.cpp release (b9590), verifies its sha256, rewrites rpaths, ad-hoc re-signs, and installs the binary plus its dylib closure; wired as externalBin + bundled Frameworks in tauri.conf.json, cached in CI, gitignored.

How it works

The runner owns the process lifecycle through message passing; subsystems hold an EngineHandle and await ensure_loaded(target). Downloads write to tmp/<sha256>.partial, resume via Range, verify the digest, then move into the blob store and insert a manifest row; blobs are refcounted across rows so shared files (for example a vision projector) survive deleting one model. The sha256 check is an integrity check (truncation, bit rot, resume corruption); provenance comes from downloading only pinned repo revisions.

Testing

bun run test:all:coverage and bun run validate-build both exit 0 at the branch head (100% coverage gates, frontend and backend). The runner is tested under paused tokio time (idle, restart, crash recovery, single-process invariant); the downloader against local mock servers (resume, cancel during stall, checksum failure, disk-full classification).