3. Add Vectors

Use matching beta versions so both crates use the same token types:

[dependencies]
archmage = "=0.9.31-beta.2"
magetypes = "=0.9.31-beta.2"

define(f32x8) gives each generated body an eight-lane vector alias specialized for its tier. Obtain that tier's proof locally with Token::from_context(). The caller still passes only data.

Legacy equivalent: #[magetypes(define(f32x8), ...)] supplies the aliases and variants; an incant! dispatcher provides the ordinary calling interface.

#![forbid(unsafe_code)]
use archmage::prelude::*;

#[attune(dispatch, define(f32x8))]
pub fn scale(values: &mut [f32], factor: f32) {
    let token = Token::from_context();
    let multiplier = f32x8::splat_t(token, factor);
    let (chunks, tail) = f32x8::partition_slice_mut_t(token, values);

    for chunk in chunks {
        (f32x8::load_t(token, chunk) * multiplier).store(chunk);
    }
    for value in tail {
        *value *= factor;
    }
}

// Exercise an empty slice, a short slice, an exact chunk, and a chunk with a tail.
for len in [0, 3, 8, 11] {
    let mut values = vec![2.0; len];
    scale(&mut values, 0.5);
    assert_eq!(values, vec![1.0; len]);
}

Token is a placeholder that the generated family replaces with its concrete tier. from_context() uses the enclosing feature guarantee without another CPU probe. Magetypes' _t(token, ...) constructors still require that proof in this beta; short constructor names have not become token-free.

The slice partition gives you complete eight-element arrays and a remaining slice. Process both. f32x8 always has eight logical lanes here, including on the scalar path; it does not automatically widen when the CPU supports more lanes.

For an image with row padding, call the kernel on each row's active slice. Keep stride and padding handling in the row traversal rather than treating the whole allocation as contiguous pixels.

Continue when you need more

• Memory operations: arrays, chunks, and bounds.
• Numerical precision: choose an operation's numerical contract.
• Full attunement reference: explicit AVX-512 gates, generic signatures, proof wrappers, and output visibility.
• Vector expansion: the concrete aliases and feature bodies generated for each target.

The shared Magetypes chapters include examples using the existing spellcraft attributes. The vector operations are the same; the attuned kernel above is the starting pattern for this path.