v0.4

Added

• Zstd.versionNumber() — the linked zstd version as a single integer
  (MAJOR * 10000 + MINOR * 100 + PATCH, e.g. 10507 for 1.5.7), for
  programmatic version checks alongside version().

Changed

• ZstdSkippableContent is now a true immutable value: it defensively copies its
  bytes on the way in and out, and compares by content
  (equals / hashCode / toString over the payload, not array identity).
• Public methods fail fast with a named NullPointerException on null byte[],
  dictionary, or sample arguments, instead of an opaque failure deep in native
  code. Streams are documented as not thread-safe; digested dictionaries
  (ZstdCompressDict / ZstdDecompressDict) as immutable and safe to share.

Fixed

• A streaming wrapper that failed partway through construction (e.g. an invalid
  parameter or dictionary) leaked the native context. The context is now freed
  on every constructor error path.

Security

• The bundled library is extracted into a directory created owner-only
  (rwx------) atomically at creation, not just by default. The third-party
  setup-zig CI action is pinned to a full commit SHA.

v0.3

Changed

• zstd-platform is now an ordinary (empty) jar instead of a pom aggregator:
  depend on it like any other artifact — drop the <type>pom</type> you needed
  before. It still transitively pulls the bindings plus all six native libraries.

Security

• Native loading is bundled-only. Removed the -Dzstd.lib.path override —
  loading a caller-supplied native library is arbitrary native code execution in
  the JVM, so the loader now trusts only the artifact bundled on the classpath.
  The bundled library is extracted into a private, owner-only temp directory
  (closing a swap/symlink window in the shared temp root). To run a self-built
  libzstd, rebuild it into the native resource jar (see docs/how-to.md).

Fixed

• ZstdSkippableContent now compares by content: equals / hashCode /
  toString consider the payload bytes instead of array identity.

v0.2

Added

• zstd-platform aggregator artifact: one dependency pulls the bindings plus
  every platform's native library, so a build runs on any OS/arch without
  choosing a classifier.
• MemorySegment constructors on ZstdCompressDict / ZstdDecompressDict to
  digest a dictionary straight from off-heap memory (e.g. an mmap slice) with no
  heap byte[] copy.

v0.1

First release. Java 25 Foreign Function & Memory (FFM) bindings for
Zstandard 1.5.7, built hermetically from
vendored source with zig cc (no JNI, no prebuilt binaries). 68 of the public
zstd symbols are bound; see docs/supported.md.

Added

• One-shot compression/decompression over byte[] and zero-copy MemorySegment
  (Zstd, ZstdCompressCtx, ZstdDecompressCtx).
• Dictionaries: training (ZDICT_trainFromBuffer, COVER / fast-COVER optimisers,
  finalizeDictionary), digested ZstdCompressDict / ZstdDecompressDict,
  dictionary ids and header size.
• Streaming: ZstdOutputStream / ZstdInputStream (java.io) and a zero-copy
  MemorySegment driver (ZstdCompressStream / ZstdDecompressStream), with
  dictionaries, pledgedSrcSize, and live progress().
• All advanced parameters (ZstdCompressParameter / ZstdDecompressParameter)
  with bounds queries; checksum, long-distance matching, window log, etc.
• Frame inspection (ZstdFrame): header, content/compressed size, dictionary id,
  skippable frames.
• Typed errors (ZstdException.code() / ZstdErrorCode) and memory accounting
  (sizeOf(), Zstd.estimate*Size).
• Native artifacts for macOS, Linux and Windows on x86_64 and aarch64,
  cross-compiled from a single host with zig cc.
• Format-compatibility tests against the reference zstd-jni binding.