3.7.2

Changes

• Parse results are now deterministic across repeated parses, processes
  and PYTHONHASHSEED values (#66). Previously, when a parse forest
  contained subtrees with exactly equal reduction scores, the tie was
  effectively broken by memory allocation addresses: the Earley state hash
  in the C++ parser core included raw pointer values, which perturbed the
  order in which derivation families were assembled in the packed parse
  forest, and thereby which equally-scored tree the reducer returned.
  The state hash is now content-based, and grammar terminals/nonterminals
  are hashed by stable sequence numbers instead of id() on the Python
  side. Note that as a consequence, the single tree returned for an
  ambiguous sentence with tied scores may differ from the one that a
  previous version happened to return - but it is now always the same one.

• No changes to the public API.

3.7.1

Changes

• CPython wheels are now correctly tagged as abi3 (cp310-abi3), so a
  single binary wheel serves all CPython versions >= 3.10. The extension
  module has always been compiled against the stable ABI, but the wheels
  were previously tagged for the exact build version only (e.g.
  cp310-cp310), which meant that users on other CPython versions
  silently built the package from source on installation. With this fix,
  all CPython 3.10+ users get fast binary installs. (PyPy wheels remain
  version-specific, as PyPy has no stable ABI.)

• No functional changes to the parser or its API; see the
  3.7.0 release notes
  for the major changes in the 3.7 series.

3.7.0

Changes

• Native (C++) token/terminal matching — parsing is up to 2.2x faster.
  The majority of token/terminal match decisions are now made natively in the
  C++ Earley core instead of via Python callbacks, roughly doubling parsing
  speed on fresh text (~89% of matching callbacks eliminated). Matching
  results are bit-exact with the previous Python matcher, verified by a
  query-level parity mode (GREYNIR_MATCHING_PARITY=1) and an extensive
  test corpus. Derived parsers that override token wrapping (such as
  GreynirCorrect) automatically continue to use Python matching.

• Python >= 3.10 is now required. Python 3.9 is end-of-life; users on
  3.9 should stay on GreynirEngine 3.6.x. Binary wheels are now built as
  cp310 abi3 for CPython, plus PyPy 3.11 wheels.

• Robust interprocess locking and atomic grammar writes. The binary
  grammar file is written atomically (temp file + rename), so an
  interrupted process can never leave a truncated grammar behind. The
  hand-rolled lock in glock.py has been replaced by the cross-platform
  filelock package; the lock file now lives next to the binary grammar
  instead of in /tmp, lock acquisition times out with a clear error
  instead of hanging, and no lock is taken at all when the binary grammar
  is already up to date.

• Hardened parser core. The C++ core now validates the binary grammar
  file fully on load, catches out-of-memory conditions at the C ABI
  boundary instead of crashing, uses race-free diagnostic counters, and
  bounds the token matching cache (~125 MB) in long-running processes.

• Greynir() now raises TypeError on unrecognized keyword options
  instead of silently ignoring them. Note that e.g. max_sent_tokens is a
  parameter of the parse methods, not of the constructor.

• Modernized tooling. The project now uses uv with a committed
  lockfile and PEP 735 dependency groups; PEP 639 license metadata;
  mypy-clean type checking as a CI gate; and CI testing on Windows and
  macOS in addition to Linux (CPython 3.10-3.14 and PyPy 3.11).

3.6.2

Changes

• Fix undefined symbol: __gxx_personality_v0 on Linux CPython builds.
  The C++ extension is now linked correctly so importing the compiled parser
  no longer fails with a missing C++ runtime symbol on Linux CPython.

• Require islenska>=1.3.2. This pulls in a BÍN-lookup fix where a word
  that is a whole BÍN entry but appears in a capitalized surface form
  (e.g. an address) was mis-resolved as a synthetic compound. As a result,
  NounPhrase declension is corrected — e.g. NounPhrase("Æðarvarp 17")
  now yields Æðarvarp 17 instead of Æðar-varp 17.

Version 3.6.1

What's New

• abi3 wheel support: Single CPython wheel now works across Python 3.9-3.14+, significantly reducing release size and build time
• Smaller releases: Reduced from ~15 CPython wheels to just 3 abi3 wheels (one per platform)

Compatibility

• CPython 3.9 - 3.14+
• PyPy 3.11
• Linux (x86_64), macOS (x86_64, arm64), Windows (AMD64)

Version 3.6.0

What's Changed

Breaking Changes

• Removed deprecated KLUDGY_ORDINALS_PASS_THROUGH, KLUDGY_ORDINALS_MODIFY, and KLUDGY_ORDINALS_TRANSLATE exports (these were removed from tokenizer 3.6.0)
• The handle_kludgy_ordinals parameter to Greynir() is no longer supported

Dependencies

• Now requires tokenizer>=3.6.0

CI/CD Improvements

• Added automatic PyPI publishing via trusted publishing (OIDC)
• Added source distribution (sdist) to releases
• Updated wheel builds: CPython 3.9-3.13, PyPy 3.11
• Platforms: Linux (manylinux x86_64), macOS (x86_64 + arm64), Windows (AMD64)

Other

• Updated copyright year to 2025

Version 3.5.8

• Various minor fixes
• Project modernized and migrated to pyproject.toml
• Explicit support for Python 3.14

Full Changelog: 3.5.7...3.5.8-4

Version 3.5.7

• Now explicitly supports Python 3.13 and requires Python 3.9+
• Improved documentation

Full Changelog: 3.5.6...3.5.7

Version 3.5.6

• Minor bug fixes and tweaks

Full Changelog: 3.5.5...3.5.6

Version 3.5.5

• Fixed critical parsing bug introduced in version 3.5.4
• Now explicitly supporting Python 3.12

Full Changelog: 3.5.4...3.5.5