Function cuda_free_memory_bytes 

pub fn cuda_free_memory_bytes() -> Option<usize>

Query GPU free memory without requiring a GpuBackend reference. Safe to call from any thread that shares the CUDA context.

Applies the same rule as AtlasCudaBackend’s free_memory: host MemAvailable stands in for the driver’s figure ONLY on an INTEGRATED GPU (GB10 and friends), where cuMemGetInfo reports Linux MemFree and so omits reclaimable buff/cache. On a discrete card host RAM is a different pool and the substitution reports many times the card’s capacity — which pinned the watchdog’s reading near total host RAM, so it could never cross its threshold, and put the same fiction on the TUI’s memory gauge.