JPEG implementation and vectorization

JPEG runtime code remains safe Rust. The production kernel boundary is src/codecs/jpeg/kernels.rs, with exact scalar reference behavior and safe wide arithmetic. The historical jpeg-wide-color feature is retained as a compatibility alias for jpeg; the JPEG feature uses safe SIMD kernels on the encoder and decoder paths where they are admitted.

The direct baseline and progressive reconstruction paths use the portable wide kernels on AArch64 and x86_64. The progressive path retains its checked scalar fallback for inputs outside the parser-proven coefficient range. The x86_64 build uses its SSE2 baseline; the benchmark lane also builds an AVX2 variant. Production dispatch does not detect AVX2 at runtime, so a binary compiled with +avx2 requires an AVX2-capable deployment CPU. Use the paired same-host measurements before claiming that either x86 build is faster.

The baseline 4:2:0 encoder keeps four FDCT coefficients in packed lanes through quantization. Safe u32x4::widening_mul computes their reciprocal products together; the possible one-step quotient correction stays exact as a 0-or-1 subtraction. Keep this path byte-identical across the public Pillow JPEG encode matrix.

The token-aware generic encoder also transforms four adjacent blocks with the same safe FDCT kernel. It preserves row and per-block cancellation polls; a poll can be delayed by at most one four-block transform batch. Tail lanes repeat the final source block and are discarded after the transform.

Evaluate a change using the complete public API benchmark, including odd dimensions, quality, chroma sampling, progressive/optimized modes, and grayscale/CMYK. A faster inner loop can lose end-to-end after allocation, conversion, or routing costs. Keep context/output lifetime inside both measured implementations' boundaries.

Do not import architecture-specific unsafe intrinsics, unchecked pointer loads, uninitialized output buffers, or fixture-specific fast paths to claim parity. Preserve exact fixed-point rounding, saturation, and byte output with maintained Pillow cases. Document the earliest reference divergence beside a subtle fix.

Historical rejected candidates and their local measurement diaries remain in Git history. Current evidence comes from the maintained same-host benchmark, source-bound reports, and the full fixture matrix.