The capture daemon read the whole clipboard payload into memory and handed
it to Pillow, zbarimg and tesseract with no limit. Payloads now stream
through a capped reader (32 MiB images, 4 MiB text, 5 s deadline) and are
dropped when exceeded; image dimensions come from the container header
without decoding, and QR/OCR only run under 40 megapixels. The OCR backfill
script applies the same pixel guard. Limits are overridable via
CLIPBOARD_MAX_IMAGE_BYTES, CLIPBOARD_MAX_TEXT_BYTES, CLIPBOARD_MAX_PARSE_PIXELS.