chore: release v0.6.7

[MagicalTux]

🤖 New release

• compcol: 0.6.6 -> 0.6.7 (✓ API compatible changes)

Changelog

0.6.7 - 2026-06-30

Added

• (zstd) validate Content_Checksum frames (XXH64)

Fixed

• (decoder) report OutputFull when the output buffer is full

Other

• (lzma2/xz) linear-time match finder + faster optimal parse

Added

• (zstd) decode and validate frames carrying a Content_Checksum (the zstd
  CLI writes one by default). Previously any such frame was refused with
  Unsupported, so default zstd output only decoded with --no-check. Adds a
  streaming XXH64 implementation; the decompressed output is hashed and checked
  against the 4-byte frame trailer, reporting ChecksumMismatch on corruption.

Fixed

• (decoder bridge) a decoder that buffers a whole block internally (notably
  bzip2) could fail a naive decode loop with UnexpectedEnd. When the
  caller's output buffer filled mid-block, the RawDecoder→Decoder bridge
  reported InputEmpty (because the decoder had already absorbed all the input)
  instead of OutputFull; a loop that stops on InputEmpty then called
  finish on a half-drained stream and got UnexpectedEnd. The bridge now
  returns OutputFull whenever the output buffer is full, which is always the
  correct "drain and call again" signal. Affected bzip2 round-trips whenever a
  decoded block was larger than the output buffer.
• (lzma2/xz) eliminated quadratic encode time on incompressible/low-match
  input. The match finder's hash head table was a fixed 64 Ki buckets, so as the
  input grew the per-bucket chains lengthened and every probe walked work that
  scaled with the input — xz encode of 4 MiB of random data took ~6.7 s and
  kept worsening. The head table is now sized to the match-finder window (like
  liblzma sizes its hash to the dictionary), so chains stay O(1) and encode is
  linear. Output is byte-for-byte unchanged.

Changed

• (lzma2/xz) faster optimal-parse encoder, output unchanged: length-symbol
  prices are cached per pos_state and refreshed periodically (instead of an
  8-bit bittree walk per length per position), the new-match distance price is
  computed once per dist-state band rather than per length, and match-length
  comparison runs eight bytes at a time. Net: ~3× fewer instructions on
  natural-language text, ~4× on long-run data, ~1.6× on mixed source code, with
  identical compressed output.

---

This PR was generated with release-plz.