Function background_load_allowed 

pub fn background_load_allowed() -> bool

Whether this process may load the ONNX model on a detached background thread (#519).

ONNX Runtime registers its op schemas in global C++ static state while a model is loading. If a detached loader thread is still mid-load when the process returns from main, it races libonnxruntime’s static-destructor teardown — a use-after-free SIGSEGV inside onnx::OpSchema on an ORT worker thread. The shipped lean-ctx daemon/MCP server is long-lived, so its warmup always finishes well before exit; short-lived processes (cargo test/ bench/doctest binaries, build-time generators) can exit mid-load, so we refuse the background spawn for them. Their semantic features simply stay cold — blocking, on-thread loads still work and always complete before exit, which is race-free.

Note: blocking shared_engine loads are intentionally NOT gated — they finish on the caller’s thread before the process exits, so no worker is ever active during teardown.