GPU final: model resident on cuda (no per-call transfer) + 240s budget + one-shot — fixes empty-output timeout

REAL MODEL on reliable CPU: llama.cpp 4B Nemotron, one-shot, compact prompt; cpu-basic hardware

match the working bench EXACTLY: .to(cuda) inside the GPU call every time (no caching) — fixes the persistent NVML on the live single call

ONE-SHOT plan generation for ZeroGPU: whole wish in a single @spaces.GPU call (ends the multi-call NVML crash, far faster) + deterministic town fallback so a wish never builds nothing

FIX multi-call NVML crash: eager .to(cuda) under spaces at load (manages GPU residency across a wish's many calls) — not lazy-per-call

back to zerogpu (CPU-only app is killed on ZeroGPU hardware) + retry transient NVML/CUDA failures on a fresh GPU schedule

switch live backend to llama.cpp (CPU, GBNF-constrained) — reliable, ends the ZeroGPU NVML flakiness; max threads

FIX zerogpu NVML crash: defer cuda placement into the GPU context (no startup CUDA init); reject failed wishes honestly instead of empty-granting

strip echoed READING label; fix status sky suffix accretion

gradio Server mode + browser-invoked granting (the her idiom)

transformers v5 in-tree stack

switch to 9B-v2 in-tree, kernel-free

trust_remote_code for nvidia repos

GODSEED v0