perf(text): evaluate adjacent-texel SIMD

[thejustinwalsh]

Summary

• add feature-gated scalar-tile and adjacent-texel SIMD MTSDF kernels without changing the production kernel
• extend the Node and Chromium runners to compare all five candidates with exact oracle and allocation checks
• record complete Inter, browser timing, Wasm size evidence, and the rejection decision in canonical package docs
• scope experimental compiled-module bytes to the recorded host while keeping portable quality evidence exact everywhere
• refresh this layer's fail-closed package-size provenance

Decision

Keep the scalar kernel in production. Adjacent SIMD improved the bounded Chromium median by 0.9% (28.700 ms versus 28.965 ms), did not produce a stable Node advantage, and increased optimized Wasm by 20.7% and Brotli size by 11.4%.

Verification

• complete Inter: 2,937 slots, 2,915 generated glyphs, 22 exact rejected IDs, exact composite identity
• every host rebuilds all five variants and must retain target features, zero imports, seven native-oracle hashes, allocation counts, and steady-state memory behavior
• recorded-host module bytes are exact; foreign-host optimized/gzip/Brotli measurements must remain under per-variant ceilings and carry complete SHA-256 identities
• @pmndrs/text full tests, Rust unit tests, clippy with warnings denied, and fuzz smoke pass
• benchmark unit: 45 files / 257 tests; script tests: 3 passed
• package-size, root typecheck, lint, format, and OKF validation pass

Stack

Depends on #8. No experimental kernel or runtime toggle ships from this PR.