mirror of
https://github.com/storytold/pdfcraft.git
synced 2026-10-25 18:47:22 -06:00
skylar taylor-barrick
7dc94ce431
Review findings, each reproduced before the fix: - row_hits built each transformed row's bounding box and skipped it on a strict two-axis overlap; a 90-degree cm makes every row vertical (zero X extent), so every row was skipped and the original JPEG survived the mark untouched behind the overlay. A degenerate row now falls through to the per-pixel test; regression: a rotated scan is cleared. - a mark that clips an image's edge without covering a pixel centre left the image as-is; both the JPEG and raw paths now remove such an image whole instead (the mark touches it, fail closed). Regression: an edge-sliver mark removes the image. - identity_decode was checked against three components unconditionally, so a DeviceGray scan with an explicit /Decode [0 1] was removed whole. The check now uses the image's own component count. Regressions: gray with identity Decode is cleared; inverted Decode still fails closed. - the JPEG header's dimensions are now read before the sample allocation, so a dictionary that lies small cannot decode a huge JPEG past MAX_PIXELS; mismatches fail closed as before. 46 redact tests green; clippy/fmt clean; wasm32 check passes.