GPU tuning: @spaces.GPU duration 90->60, max_new_tokens 768->320, remove diagnostic logging

Warm VoxCPM2 at startup (under ZeroGPU emulation) + load_denoiser=False — fixes lazy-import CUDA-init error

Shim legacy generation output classes so the mamba-ssm hub kernel imports on transformers 5.x

Nemotron-3-Nano fast Mamba path: native NemotronH (transformers>=5.8) + hub kernels, drop mamba_ssm wheel (fixes NVML fork assert)

Load Nemotron on CPU, move to GPU inside @spaces.GPU (fixes NVML fork assert)

Disable Nemotron thinking (enable_thinking=False) so the dispatcher emits clean JSON

Part 2: @app .api game endpoints + @gradio/client frontend (ZeroGPU auth-forwarded GPU calls)

Expose demo=app for gradio reload mode (it launches the demo global)

app_file=frontend/server/app.py with SPACE_ID-gated module launch, ssr_mode=False

Switch to gr.Server (Gradio Server mode) — fixes the 7860 double-bind

Deploy Mumbai Local Panic — Nemotron(ZeroGPU) dispatcher + VoxCPM2 voice