- finish_8bit/finish_deep hand back the render itself when there is no
output sharpening and no watermark. Before, every export cloned the
image (~100 MB at 24 MP) and output_sharpen built a ~290 MB float copy
of it even for SharpenFor::None, all on one thread.
- Output sharpening's box blurs run row-parallel (par_rows); each pixel
is computed as before.
- The parallel JPEG encoder converts each band's rows to YCbCr inside the
band's task instead of first filling whole-image planes (~290 MB for
4:4:4 at 24 MP, written once and read back).
- fdct_quant rounds with plain arithmetic (round_i16) instead of
f32::round, a libm call per coefficient without SSE4.1; it equals
`x.round() as i16` for all 2^32 inputs.
Output is byte-identical (pinned by encoded_bytes_are_pinned and
sharpened_pixels_are_pinned, whose hashes come from main).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Dylan M. Taylor