Summary

Platform detection functions (has_nvidia, has_gaudi, has_tenstorrent, etc.) in src/device/platform_detection.rs were re-evaluated on every view refresh cycle. On macOS this meant system_profiler SPPCIDataType — a probe that takes hundreds of milliseconds and spawns a fresh process — was running once per frame in local view mode.

Root Cause

In src/view/data_collection/local_collector.rs:565, update_notifications is called from every data collection cycle and evaluates has_nvidia() first in a short-circuit chain:

if has_nvidia()
    && let Some(nvml_message) = get_nvml_status_message()
    && !state.nvml_notification_shown

Because has_nvidia() is the first term, the nvml_notification_shown guard never short-circuits the probe. On macOS the function shells out to system_profiler SPPCIDataType on every call.

Fix

Wrap each platform detection function in a process-global OnceLock, so the underlying probe runs at most once per process:

• has_nvidia, has_amd, has_furiosa, has_tenstorrent, has_rebellions, has_google_tpu, has_gaudi
• is_jetson, is_apple_silicon

Hardware presence does not change at runtime, so caching is semantically correct. All call sites in reader_factory, local_collector, main, and client automatically benefit without modification.

Also collapses nested if blocks flagged by clippy::collapsible_if in macOS-specific device readers (cpu_macos.rs, macos_native/*.rs, apple_silicon_native.rs).

Test plan

• cargo build succeeds
• cargo clippy clean
• cargo test passes (5 + 17 + 19 tests)
• Manual verification on macOS that system_profiler is no longer spawned on every view refresh cycle