Adoptable techniques from the fearless_simd review (2026-09): swizzle_dyn, safe token materialization, formulation wins, docs batch

A model-driven review of linebender/fearless_simd (0.7.0, changelog 0.2.0→0.7.0, recent PRs/issues, and the related blog posts) against archmage/magetypes, looking for improvements worth adopting. fearless_simd is well-designed, fast-moving work and several of its recent changes carry lessons we can reuse — full report on the review box at `~/tmp/fearless-simd-review-2026-09-04.md`. Ranked adoptables:

1. **f32→i32 conversion contract** — split out as #80 (verified live divergence here).
2. **Safe token materialization**: their `assume_supported()` is a `#[target_feature(enable=…)] const fn` constructor — callable *safely* from matching feature contexts, `unsafe` elsewhere (their #293). This answers our open 'rite token forging' design question (CLAUDE.md Open Questions) with zero user-visible unsafe and no new safety model: the `#[target_feature]` context is the proof. Landing site: `xtask/src/token_gen.rs` → generated token impls.
3. **Dynamic byte swizzle family** (`swizzle_dyn` / `_precise` / `_within_blocks`, their #266/#276/#304/#322/#362): magetypes currently has no pshufb/tbl-class op at all — the workhorse for palette expansion, format shuffles, and byte transforms. Their two-tier out-of-range contract (relaxed vs zeroing) is a clean answer to the ISA divergence, and the AVX2 formulation they landed (control-derived blend mask) is documented with measurements. Landing site: new u8xN backend methods via xtask, with lane-order-pinned tests.
4. **Formulation wins worth an llvm-mca pass over our backends** (their #350/#291/#343/#344/#279): 8-bit multiplies via PMADDUBSW (we currently expose no 8-bit mul), a cheaper NEON bitmask than our per-lane `vgetq_lane` packing (`impls/arm_neon.rs:1429/1695`), polyfill `all_true(a,b)` as `reduce(and(a,b))` rather than two reductions (`impls/arm_neon.rs:1679`), plus interleave and AVX-512 byte-shift lowerings.
5. **Docs/table batch**: note the std `mul_add` subnormal fallback bug (compiler-builtins#1262) next to our fusion-contract docs; a '`round` ≠ Rust scalar `round`' warning; IEEE-vs-native min/max naming (`min_ieee`/`max_ieee`); never implement trunc/fract via int roundtrip (their #365: `a-(f32)(i32)a` breaks at ≥2³¹); and an `incant!`-adjacent gotcha note about `return`/`?` inside dispatched blocks.

Also worth knowing: their build-time lesson (derive shared-shape op impls from common generic code instead of per-backend×width emission; −25% metadata, meaningful for our 84-file generation), a `#[doc(hidden)]` macro-surface soundness-hardening audit prompt for archmage-macros (their #363), and one capability we lack by decision: 32-bit `target_arch = "x86"` support (our tokens stub out on i686 — worth re-confirming that's intentional given i686 is a primary CI target).

Things we deliberately should NOT copy (and why) are in the report — mostly places where our existing design (registry-driven tiers, token caching, uniform out-of-range shift contracts, no-silent-test-skips) already sits where their issue tracker is pointing.

Credit where due: the review leaned on their public issue discussions and Shnatsel's writeups, which document these techniques unusually well.