Use dynamic model discovery and hardware-aware recommendations

[SnowCheetos]

Context

Common model options are currently static. A more dynamic approach could query Hugging Face model data and tailor recommendations based on local hardware, so users with more capable machines can be guided toward larger or higher-quality models.

Goals

• Discover common model options dynamically, likely via the Hugging Face API.
• Inspect local hardware signals where available: OS, architecture, memory, GPU/backend support, and maybe disk space.
• Recommend model sizes and quantizations based on the user’s machine.
• Preserve an offline/static fallback so configuration still works without network access.

Acceptance criteria

• Model selection works with and without network access.
• Recommendations explain the chosen model class briefly in CLI output.
• Hardware probing is best-effort and does not fail configuration when a signal is unavailable.
• Tests cover dynamic data parsing and fallback behavior.