Lossy JPEG COGs for visual imagery

Context

Lossless ZSTD can expand an already JPEG-compressed ortho by 10x. The uploaded file is retained as the unmodified original, while the COG is the serving copy.

WEBP q90 is 10-20% smaller than JPEG q90, but several GDAL builds cannot read WEBP-in-TIFF, including QGIS on macOS (#342). Legacy OAM, Maxar, NOAA, and Vantor already use YCbCr JPEG (#337).

Decision

Write visual uint8 RGB(A) as YCbCr JPEG quality 90; keep other products lossless ZSTD-12. Transparency becomes a per-dataset mask band. Verify georeferencing, the validity mask, sampled pixel error, and COG layout.

Fix codec and quality in code to match validate's measured disk allowance. Higher quality needs more space; convert-only settings would be invisible to validate.

Budget 0.9x decoded size for lossy output and 1.5x for lossless. JPEG q90 peaked at 0.72-0.83x in measurements (#365). Keep GDAL's overview temp lossless ZSTD: lossy temp files halve peak disk use but reduce overview quality to 36 dB at q90.

Record the lossy codec in lineage and the visual asset title. Existing items are not converted again.

Consequences

Visual COGs are smaller and avoid WEBP compatibility issues, but recompression can reduce quality. Readers that ignore mask bands (geotiff.js) show black collars. Originals stay byte-identical; non-visual products stay lossless.