v3.17.0 — Smarter model pulling + Ollama RAM/server cleanup

Make pulling any model "just work", with a much better pull UX — plus Ollama RAM/server cleanup on exit.

Added

• HuggingFace GGUF fallback — when a model isn't in Ollama's registry (or cloud), Kodiqa searches HuggingFace for a community GGUF build, lists the quant levels with sizes to choose from, and installs it via ollama pull hf.co/<repo>:<quant>. Escalation per model: Ollama registry → <name>:cloud → HuggingFace.
• Live download progress — pulls stream Ollama's real progress bar instead of a static "Pulling…" line.
• Download sizes up front — the startup list and /pull show each model's size (concurrent registry lookup, no downloads); the confirm shows a total. Cloud models are labelled ☁ cloud, unknowns size ?. Models without a latest tag are resolved via the library tags page, so non-standard cloud tags (mistral-large-3:675b-cloud) are labelled cloud and sized-only models (granite4.1-guardian:8b) show a real size and pull using the resolved tag.
• Paginated model list — up to ~234 models, 100/page, next/prev, global numbers, lazy per-page sizing.
• Pick-again loop — n at the confirm returns to the picker; skip cancels.
• Cancel a running pull — Esc (confirm) or Ctrl+C (immediate); the partial download is cached so re-pulling resumes.

Fixed

• Ollama no longer leaves a 10GB+ model resident in RAM after you quit — on exit Kodiqa unloads loaded models (keep_alive=0). Disable with unload_ollama_on_exit: false.
• A Kodiqa-started Ollama server is now stopped even across sessions (via a PID file), while never touching a GUI-app or user-started server.

527 tests pass, ruff clean.

🤖 Generated with Claude Code