Replace SIXEL retention and the per-client cache with the global indexed-image cache from image_merge. Adapt its backend callbacks for Kitty and test palette quantization through both input protocols.
Keep retained indexed SIXEL data and the bounded per-terminal output cache.
Use the cleaned-up decoder and RGBA encoder from image_merge, including
parser memory limits, pixel-size handling and references for pending output.
Match retained SIXEL scaling to the shared image cell footprint and emit
black definitions for undefined colour registers.
Keep Kitty input, output, layers and placeholders. Adapt backend selection,
flags, drawing and cleanup to the callbacks and Iscrl capability. Keep the
pane-status repaint with the requested XXX comment.
Replace the elaborate fallback with image_merge's implementation, removing
its caches, sample structures, quadrant/sextant features and ACS additions.
Keep the image_merge regressions and adapt the Kitty tests to the simpler
fallback. Use skip status 101 and disable Python bytecode for image tests.
Small changes still outside image_merge:
- Guard zero-height synchronized redraws before bit_nset.
- Select the image backend when the tty opens.
- Enable imagescroll by default for Kitty and Ghostty.
- Keep the explanatory manual-window pixel-size comment (comment only).
- Keep the additional SIXEL fixtures and visual test tools.
Validation: image-enabled and disabled builds; 31 image and 13 related
regressions pass normally and under ASAN. All image tests skip without
images. Focused palette, parser, scaling, transparency and Kitty layer
tests also pass after the retained-output adjustments. Diff checks clean.
