Wikipedia says "after bit stuffing, the data string is padded to the next
codeword boundary by appending 1 bits. If this would result in a code word
of all ones, the last bit is changed to zero", we were missing the second
part here.

While this difference wouldn't confuse the decoder normally, it is also
fed into the Reed-Solomon encoder, and thus spread out and multiplied,
resulting in a non-decodeable image.

This is sufficiently unlikely (needs a content ending with a binary
content with enough bits set, and unfortunate alignment of the bit
stream), which is why this went unnoticed for quite a while.