v2.3.0 — TUI width guard + hf-mem 0.5.5

✨ Features

• TUI width guard. When the terminal narrows below LlmuxApp.MIN_WIDTH = 80, a TooNarrowScreen overlay is pushed so modals, forms, and the dashboard no longer clip horizontally. Auto-dismisses the moment the terminal is widened back — nvitop-style behavior. (#55)

⬆️ Dependencies

• hf-mem 0.5.2 → 0.5.5. Lock-only bump, no call-site changes. Picks up:
  • KV-cache dtype resolution for compressed-tensors models that don't declare kv_cache_scheme (e.g. Kimi K2.5)
  • Sliding-window layer estimation fix → more accurate KV-cache numbers for SWA models
  • Missing dtype / quantization_config parsing fix
  • New: modelopt quant_method, F8_E8M0 dtype, experimental MoE breakdown for safetensors (relevant for openai/gpt-oss-120b-class models shown in dashboard)
  • Upstream: https://github.com/alvarobartt/hf-mem/releases (#55)

📝 Docs

• README: dropped the obsolete "Credits" block pointing at the now-superseded vllm-compose / llamacpp-compose repos.

Changelog

• #55 ✨ feat(tui): width guard against terminals narrower than MIN_WIDTH=80
• #55 ⬆️ chore(deps): bump hf-mem 0.5.2 → 0.5.5
• #56 📝 docs(readme): drop superseded project credits
• #56 🔖 chore: bump version to 2.3.0

Full diff: v2.2.1...v2.3.0