NEWS

Release notes for xpar. See ChangeLog for source-level changes.

v2.0 (26-08-2026)
- Version 2 uses a new, incompatible container and has no v1 decoder.
  It identifies and refuses v1 files. Decode with xpar 1.x, then protect
  the result with xpar 2; there is no in-place upgrade.
- Verbs replace mode flags. The v1 flags '-J',
  '-W', '-L', '-s', '-t', '--interlacing', '-H / --integrity' and
  '--auth' are gone, as are v1 shard files. V1 scripts require changes.
- doc/xpar-format.tex is the normative format specification.
- xpar now protects file sets and directory trees. Manifests record
  names, sizes, modes, four timestamps, POSIX ownership, extended
  attributes, hard links and symbolic links; '--preserve' controls
  recording and extraction.
- Two Reed-Solomon layers: an outer cell erasure code limits sector
  damage to one (slice, column) cell, while inner armour corrects bit
  errors within slices. '--cell', '--burst' and '--depth' configure
  erasure units and interleaving.
- FFT and systematic Cauchy-matrix outer codecs support GF(2^8) and
  GF(2^16). Selection is automatic or fixed with '--codec' and '--field'.
- In-place repair preserves names, inodes and hard links while rewriting
  only damaged cells. A durable undo journal precedes destructive writes;
  'xpar undo' replays journals left by interruption. Other policies use
  '--backup', '--to', '--dry-run' and '--paranoid'.
- 'consolidate --replace' and 'prune' write a durable maintenance journal,
  'base.xparmaint', before renaming a chain aside: an interrupted run is
  reported as such rather than as a missing set, and 'xpar repair'
  completes or rolls it back.
- A read the host refuses is an I/O error (status 5) and never damage: a
  volume that fails under its mapping names its file instead of dying as
  a defect, and repair never writes over a file it could not read.
- 'verify' and 'repair' can resynchronise after inserted or deleted bytes
  with '--resync'.
- Layouts: sidecar (default) leaves data in place; split writes bare data
  volumes that concatenate exactly; armoured writes one self-contained,
  inner-coded file with a replicated prologue.
- '--armour=all' now works in the sidecar and split layouts, wrapping each
  slice-table and recovery-slice packet in an inner code of its own so bit
  rot in a recovery volume costs parity symbols, not whole slices.
- Generations: 'add' records changes while inheriting unchanged data;
  'prune' removes unneeded generations; 'consolidate' collapses a chain;
  'info --deps' shows dependencies.
- '--dedup' supports whole-file and content-defined chunk deduplication.
- New verbs: 'scrub' reports inner-code corrections; 'recover' rebuilds
  a lost volume; 'addrecovery' adds redundancy; 'extract' unpacks owned
  layouts; 'explain' prints a dd(1) recovery recipe; 'recover-prologue'
  reconstructs an armoured prologue; 'benchmark' checks and times SIMD.
- Integrity uses BLAKE3-256 content hashes, 8- or 16-byte BLAKE3 slice
  tags, and CRC32C per slice and cell. '--auth-key=FILE' uses a keyed
  BLAKE3 MAC; authenticated operations require the correct key.
- Packet checksums and strict length validation reject truncated or
  corrupted containers before parsing their bodies.
- A planner enforces the '-m' working-set limit and reports parameters
  that would fit when no plan does.
- Output supports JSON Lines ('--json'), progress and ETA ('--progress'),
  colour control ('--color') and reproducible sets ('--reproducible').
  Exit codes 0-8 distinguish clean, repairable, unrepairable, not found,
  usage, I/O, authentication, no-plan and internal errors.
- SIMD support: SSSE3, SSE4.2, AVX2, GFNI, GFNI-512 and VBMI
  on x86; NEON, PMULL and SVE on ARM; VSX on POWER; shuffle and clmul
  RISC-V vectors. Runtime selection can be fixed with '--simd'. Portable
  threading uses '-j'; Linux may use dependency-free io_uring batching.
- Ports: POSIX, Windows NT, i486 Windows 95 and DJGPP/MS-DOS. The format
  is identical across platforms. CI covers x86-64, aarch64, ppc64le,
  riscv64, big-endian s390x, armhf, macOS, Windows and DOS with sanitizers.
- Tests include unit, codec, container, damage, fault and safety suites at
  three depths: 'make check', 'make check-full' and 'make check-torture'.

v1.1 (20-04-2026)
- The format is unchanged and fully compatible with v1.0 in both
  directions.
- '--progress' reports throughput and ETA during joint, systematic,
  Leopard-sharded and Vandermonde-sharded encoding and decoding.
- Optional Linux io_uring batches parallel I/O in joint and sharded
  pipelines. '--enable-liburing' enables it; Linux auto-detects liburing.
  Other builds retain the portable POSIX/Win32 backend.
- Joint encoding now requires '-f / --force' to overwrite output.
- Keyed MAC and tag comparisons are constant-time.
- Sharded decoders now share bounded consensus voting and stricter
  layout validation, including truncated or mismatched final shards.
- Leopard fixes single-data-shard decoding and verifies durable output
  before success. Vandermonde fixes an error-path buffer leak. Shard-count
  errors are clearer.
- Hardened yarg allocation and formatting paths. Reworked the legacy
  Windows command-line parser for safe growth and OOM cleanup.
- POSIX descriptors use FD_CLOEXEC; I/O retries short transfers and EINTR.
  DJGPP gains a higher-precision PIT timer and <io.h> I/O. DOS releases use
  LTO, function sections and section garbage collection.
- Release tarballs have reproducible mtimes. CI adds Win95 and DOS release
  workflows.

v1.0 (16-04-2026)
- Format bump to v1.0. v0.x archives are rejected and must be re-encoded;
  future v1.x releases remain decodable by v1.0+ tools.
- Joint headers record total input size and lace sequence numbers,
  rejecting archives truncated at a lace boundary or with reordered laces.
- Sharded-header CRC32C covers the version, shard count and number, total
  size and body, detecting field corruption and swaps between shards.
- '-H / --integrity' selects hardware-accelerated 32-bit CRC32C tags
  (default) or 128-bit BLAKE2b tags in joint, systematic, Vandermonde and
  Leopard modes.
- '--auth=<keyfile>' accepts a 1-64 byte key and uses a BLAKE2b-128 MAC,
  implying '-H blake2b'. Decoding rejects wrong, missing or spurious keys.
- '-s / --systematic' writes a parity-only .xpa, about 18% of the input,
  while leaving the original unchanged. Decoding combines both files and
  writes corrected data to stdout. It is incompatible with interlacing.
- '-t / --test' runs correction without writing for '-J', '-Js', '-W' and
  '-L'. It reports unrecoverable blocks, shard shortfalls, tag mismatches
  and archives that require correction.
- BLAKE2b uses runtime-dispatched SSE4.1, AVX2 or AArch64 sse2neon kernels,
  with a portable fallback. '--disable-blake2b' removes BLAKE2b and MAC
  support.
- Accelerated Leopard decoding on AArch64. Fixed a missing return in the
  SSSE3 xor_mem4 path that could overflow the heap when rebuilding shards.
- Self-check covers truncation, reordering and shard-number swaps. A C
  error-injection helper removes the test suite's Python dependency.
- CI adds AddressSanitizer and UndefinedBehaviorSanitizer on Linux and
  macOS. Releases add a KERNEL32-only Win95 i686 binary and a DJGPP/MS-DOS
  binary with CWSDPMI, alongside existing Windows binaries.

v0.7 (19-09-2025)
- Rename sharded mode to Vandermonde-sharded mode ('-S' to '-W').
- Add architecture- and OS-specific build instructions.
- Include benchmarks in the repository.
- Enforce a minimum shard size.
- Improve stability when mmap is enabled.
- Add linearithmic FFT-based Reed-Solomon codecs (@catid).
- Fix a gf256mat_inv leak in Vandermonde-sharded mode.
- Update the man page.

v0.6 (10-09-2025)
- Move the project page to iczelia/xpar.
- Fix command parsing and clean up its style.

v0.5 (17-10-2024)
- Add OpenMP to sharded mode; performance remains I/O-bound.
- Switch command-line parsing to yarg and remove the getopt_long dependency.
- Freeze the v0 file format except for critical defects.

v0.4 (16-10-2024)
- Stop providing static x86_64 Linux binaries.
- Add OpenMP for joint mode with high interlacing factors.
- Add three-way saturating CRC32C for x86_64 with SSE4.2.
- Improve sharded-mode performance.
- Fix undefined integer shifts in sharded mode.

v0.3 (16-10-2024)
- Improve joint encoding performance on x86_64 machines.
- Support aarch64 Linux.
- Improve cross-platform compatibility of the sharded mode.

v0.2 (15-10-2024)
- Add the xpar manual page.
- Add AArch64 code, enabled with '--enable-aarch64'.

v0.1 (14-10-2024)
- Initial release.
- Support joint and sharded error and erasure correction.
- Add x86_64 code, enabled with '--enable-x86_64'.
- Test on x86_64 Linux, x86_64 and AArch64 macOS, and x86_64 and i686
  Windows.