Friendly forks of upstream projects.

When an upstream project we depend on is unmaintained, slow to move, or simply missing something we need, we fork it here rather than vendoring a patch and forgetting about it. The fork stays honest about what it is: upstream keeps the credit, the licence stays upstream's, and every change is one we would happily see merged back.

If you can get a fix upstream, do that first. These forks exist because that was not possible or not timely — not because we think we own the code.

• xz — a fork of ulikunitz/xz, in two passes. First decoding speed: the per-operation allocations removed, a branchless range decoder, match copies by pattern doubling, and a ParallelReader that decodes the blocks of a multiblock archive concurrently. Then robustness against archives that were built to break a decoder — impossible sizes rejected, the stream index read incrementally and bounded so the backwards walk terminates, the dictionary grown on demand rather than up front, and Close able to cancel a read already blocked in a worker. Upstream appears inactive; the changes are deliberately kept in a shape that could be carried over if anyone has the time.
• erofs — a fork of erofs/go-erofs: a pure-Go library for reading and creating EROFS filesystem images through the standard fs.FS interface, no CGO. The work here is mostly about surviving images we did not write — allocations bounded by what the image can physically hold, cyclic directory graphs and oversized directory and symlink sizes rejected — alongside correctness fixes (fs.FS contract conformance, chunk extents mapped at block granularity with holes preserved, setuid/setgid/sticky bits kept on both the read and write paths) and additions to the writer.
• blake3 — a fork of lukechampine/blake3, a Go implementation of the BLAKE3 hash with AVX2/AVX-512 routines and the bao verified-streaming encoding. The changes come out of an audit of the tree: an XOF Seek to almost any buffer-unaligned offset returned output from the wrong position (latent since 2020), bao slice bounds could overflow past validation and the empty encoding was rejected, New now validates its key and size instead of panicking deep inside, and the allocation and parallelism behaviour is tuned so small XOF reads and streaming writes stop paying for work they do not use.
• grid-clock-screensaver — a fork of chrstphrknwtn/grid-clock-screensaver, a word-clock screensaver for macOS, ported to modern macOS.

Projects here are provided as-is, best-effort, without warranty.

• Contributing guide — upstream first, then sign-off (DCO), commit signing and the pull request flow.
• Security policy — never report a vulnerability in a public issue, and note that a flaw in forked code is usually upstream's to fix.
• Bugs and ideas go to the issue tracker of the repository they concern.

The sibling organization farcloser holds the tooling and libraries these are built with.