It's an interesting tale of software spelunking and edge cases, but I was wondering about this:

> Things like the uncompressed size and CRC-32 can be disregarded.

If you compute a running crc32 then you could check for that value in addition to the uncompressed size, giving you 8 bytes of precision. And if you decompress data as you go, you could check for the uncompressed size as well, giving you 12 bytes of precision. Surely that's enough to make sure you always find the real boundary of a file in the zip.