docs: brain final sign-off deltas — split version matrix (§10), quota rule E2, Post-v1 deferred items

docs: brain-ratified D8 (per-request LoRA) + final DESIGN/PLAN/README deltas

docs: reflect shipped reality — torch 2.11, per-request LoRA (D8), lazy-load (D9), adaptive residency (M2), long-form chunking (§12), sdpa-only

Security + correctness hardening (brain code review)

LoRA per-request application (ZeroGPU forks don't persist model mutations)

ZeroGPU: lazy-load models in-fork (module-level preload overran startup -> RUNTIME_ERROR)

Fix Space health check: bind 0.0.0.0 on ZeroGPU + disable Gradio SSR

Fix ZeroGPU startup: skip RAM watchdog on Space (psutil misreports in container); add packages.txt (sox/ffmpeg)

docs: add studio screenshot to README

Shorten Space short_description to <=60 chars

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