feat: faster startup, crash restart, readiness states, and robustness improvements

[AlexsJones]

What

The batch of improvements identified while reviewing the codebase after #39.

Responsiveness

• Parallel backend detection & model fetches — the 9 backend detectors (7 of which make HTTP probes with 800ms timeouts) and the 4 API model fetches now run on scoped threads. Startup and r-refresh cost one probe's latency instead of the sum; first frame renders in ~45ms locally.

Robustness

• Reader threads instead of fcntl(O_NONBLOCK) — each child's stdout/stderr is read on a dedicated thread feeding an mpsc channel. Removes the unsafe block, makes serving work on Windows (where the old code would block the UI forever), preserves the \r progress-line collapsing, and raises the log ring buffer from 200 to 1000 lines. Reader threads are joined on exit/stop so the crash popup always includes the final output.
• Graceful stop — SIGTERM with a 2s grace period before SIGKILL (unix).
• Panic hook — restores the terminal (raw mode + alternate screen) before printing the panic, so a bug can't leave the shell unusable.
• External port conflict detection — do_serve now bind-tests the port, so a port held by another process (e.g. Ollama) produces a status message instead of an instant crash popup.

UX

• Readiness states — servers probe /health once per second until ready: llama-server answers 503 while loading (verified: 000 → 503 → 200), other backends' 404 counts as up. Server cards show ◌ … loading… → ● ready; the table status column matches.
• Restart from crash popup — r relaunches the exited server with identical model/backend/settings (the launch preset now travels with the server handle).
• Host field in the serve modal (switch 0.0.0.0 ↔ 127.0.0.1), persisted by save-defaults.
• Log scrollback — j/k, Ctrl-u/d, g/G scroll the logs panel when focused; title shows (scrolled ▲N).
• Mouse capture dropped — it was enabled but unhandled, and it blocked normal terminal text selection/copy.

Polish

• GGUF header parsing — reads general.size_label, general.architecture, and <arch>.context_length from the file header (seeking past arrays, so it's a few KB read regardless of model size). Fills the Params column for files with uninformative names (e.g. gemma-3.Q8_0 → 2.5B) and the serve modal warns when ctx-size exceeds the model's trained context.
• --version / --help CLI flags.
• Deduplicated the three identical OpenAI-style model fetchers in backends.rs.

Testing

• 51 tests pass (3 new GGUF parser tests incl. truncated/corrupt headers), clippy warnings down from 39 to 27 (no new ones), cargo fmt applied.
• Exercised end-to-end in tmux against real llama-server + model: startup timing, params column, host/context fields, serve → ready transition, external kill -9 → crash popup → r restart, log scrolling, external port conflict, and clean process teardown on quit.

🤖 Generated with Claude Code