SIMD diagonal fill (NEON)

[crashfrog]

Parent

#58 (Phase 1 MVP: Evidence-Informed Alignment with Deferred Filtering)

What to build

Implement NEON SIMD-accelerated diagonal fill for WFA in phraya-align (aarch64 only). NEON is mandatory on ARM64, so no runtime detection needed - unconditionally compile NEON code on aarch64 targets. Document all unsafe blocks with SAFETY invariants.

This enables performant alignment on ARM64 servers (Graviton, Apple Silicon) without maintaining separate code paths.

Acceptance criteria

• wfa_diagonal_fill_neon(diagonal: &mut [i32]) using NEON intrinsics
• Unconditional compilation on aarch64 (no runtime detection, NEON mandatory)
• All unsafe blocks documented with SAFETY comments
• Tests: correctness (NEON result == scalar result)
• Tests: alignment correctness vs baseline scalar WFA
• Benchmark: 10kb alignment on ARM64 (scalar vs NEON, measure speedup)
• Cross-platform: no-op on non-aarch64 (x86, etc.)

Blocked by

#70 (WFA base implementation)