Pin Space torch to 2.11.0 (ZeroGPU-supported)

ZeroGPU readiness: module-level cuda preload + all-resident on 48GB card + Space metadata

Adaptive residency: evict other models when memory is tight (DESIGN §6)

LoRA live-toggle (PEFT enable/disable) + D4 UI + memory hygiene

Scaffold qwen-voice-studio: Qwen3-TTS core, Gradio studio app, design docs