A multi-megabyte base64 image src lives inline in the document's
PenNode::Image.src, so every scene rebuild clones + compares it and the
canvas decodes a full-resolution bitmap — moving a 5.4MB image lagged.
Shrink an oversized raster source the moment it enters the document, at
every host path that produces an inline data: URL into the doc src:
file-pick import / fill-image / relink (persistence_image), fetched +
searched images (image_search_session), and inline-base64 AI generations
(image_generate_host). New image_downscale module decodes via skia, fits
the longest edge to 2048px (or re-encodes a >2MB source at native size),
and emits JPEG for opaque sources / PNG for alpha — adopting the result
only when it actually shrinks. Animated GIF / WebP are detected by magic
bytes and left whole so re-encoding never drops frames.
f2f50e36cf