Platform Notes

f32x8<T> denotes eight logical lanes. The token selects the implementation, not the lane count. Use #[magetypes] to bind Token per tier and incant! to enter it, as in the first kernel.

A stronger token does not automatically implement every backend trait. Hardware feature implication and Rust trait implementation are separate facts. When using a V3 shape from a V4 context, extract token.v3() explicitly. Do not place V4 into an eight-lane tier list and assume the macro widens it.

w512 is enabled by default and provides logical 512-bit shapes. avx512 adds native AVX-512 support and implies w512. With default features disabled, request w512 explicitly if you use those shapes. Wider ARM/WASM polyfills also depend on this logical-width feature.

Generic imports are portable; explicit calls to architecture-specific functions may still require cfg guards. incant! supplies those guards. Bare architecture-bound aliases can be convenient locally, but tutorials use explicit generic types or define so the backend selection remains visible.

Use -Ctarget-cpu=native only for a known deployment CPU. summon() returns Option<Token>, not Some(true); compiled_with() is the separate query about compile-time guarantees. A global feature guarantee can eliminate detection, but is not a replacement for testing your baseline-dispatched build.