looks like low-freq AC coefficient got flipped, maybe overflow caused by overshoot-based deringing?

[jonsneyers]

(Reported by DZgas on the jxl discord)

example.png:

cjpeg -baseline -notrellis -sample 2x2 -quality 50 -tune-psnr -quant-table 1 example.png > example.jpg:

Looks to me like the AC coefficient corresponding to the lowest frequency horizontal somehow got a sign flip in that spot, possibly because of black-on-white deringing via overshoot causing an integer overflow. Adding -noovershoot produces an image that is OK, so that seems to confirm this hypothesis.