The hardest tarball?

A single GNU tar invocation has produced an archive that eight major extraction tools unpack in eight different ways, exposing unresolved ambiguities in the POSIX pax format over whether hard links may carry data. The archive, built by repeating a file so GNU tar stores the second entry as a hard link, renames that link to "café" and injects a pax header with a size=512 record—exactly one tar header—so the bytes following the link are read either as link data or as the next file's header. GNU tar, BusyBox, Python, Go and Perl ignore the size and extract file, its hard link and empty; bsdtar writes the 512 bytes into the hard link; 7-Zip, node-tar, tar-fs and Rust's tar skip the data and drop empty; Java's Commons Compress, PHP's PharData, RubyGems, R and Haskell each diverge further. The split matters because pax archives are the standard interchange format for software packages and container layers, where silent divergence can corrupt builds or installs. Watch for whether maintainers align on POSIX's "may be included" language for hard-link data blocks or add conformance tests to the format specification.