lucemia/typed-ffmpeg v4.5

on GitHub

Bug-fix release. Two separate defects made the 4.4 packages unusable for common
work: parse() could not find its cache data at all, and VideoStream.scale()
raised on every call in v7/v8/v9. Both are fixed here, so all thirteen packages
are republished.

Fixed

• VideoStream.scale() raised on every call in v7, v8 and v9. FFmpeg 7
  started printing scale's input pad and its dynamic-IO marker together, with
  the marker nested under the pad; the help parser treated them as mutually
  exclusive and discarded the pad. The generated binding then declared
  typings_input=() while still passing the stream it was called on, so
  input(...).video.scale(w=..., h=...) raised
  FFMpegValueError: Expected 0 inputs, got 1. The same bad typing broke
  parse() for any command line using scale and emitted a bogus zero-input
  ffmpeg.sources.scale(). scale was the only filter affected; v5 and v6
  predate the upstream change
• parse() raised FileNotFoundError on every install.
  pip install typed-ffmpeg[parse] installs ffmpeg-data-v9, but ffmpeg-core
  4.4 only looked for ffmpeg_data_v8 down to v5. This reproduced solely on a
  wheel install — a checkout reads packages/core's cache/list, which is not
  shipped — which is why the test suite never caught it. The lookup now
  discovers installed ffmpeg-data-v* packages by name, so a new FFmpeg major
  needs no change to it
• A data package that was installed but did not carry the requested entry
  shadowed an older one that did; the lookup now walks every installed data
  package newest-first instead of committing to the newest
• scripts/create-stubs.py listed base.py, utils/snapshot.py and
  utils/view.py as ffmpeg_core re-exports. Neither exists as a usable
  module, and the script writes unconditionally, so running it replaced working
  code with broken stubs in every version package
• Dispatching a single FFmpeg version to ci-codegen-versions.yml regenerated
  all of them: the image map lived in matrix.include keyed on ffmpeg-version,
  which is also a matrix key, so the non-matching entries were appended as extra
  jobs rather than merged
• uv.lock carried two griffelib entries after the mkdocstrings-python bump,
  which uv rejects outright; mkdocs.yml passed a mkdocstrings handler
  option that never existed and became a hard error in 2.0; the removed
  fix-encoding-pragma hook was still configured, failing every pre-commit run;
  and the filter-doc HTML fallback stopped matching after ffmpeg.org regenerated
  its pages with a newer texinfo

Added

• Regression coverage for the above: the shared suite now asserts that no filter
  reached by a stream method declares zero inputs. The only previous test
  touching .scale() was skipped unless graphviz was installed, so the most
  common filter in the library had no effective coverage
• docs/version-differences.md gains an FFmpeg 8 → 9 section, verified
  against allfilters.c / allcodecs.c / allformats.c on release/9.0.
  Three entries misattributed to 8.0 in the 7 → 8 section are corrected

Changed

• The FFmpeg version list is now derived from packages/v* rather than
  enumerated, in the scripts/ helpers (via the new scripts/_versions.py),
  .gitattributes, the ruff configuration and the codegen workflow's apply
  step. scripts/regenerate-monorepo.sh had silently skipped v9 for the whole
  FFmpeg 9 rollout, and .gitattributes left v9's generated bindings counted in
  the language statistics
• The TestPyPI smoke test installs the typed-ffmpeg meta-package instead of a
  pinned version package, so it exercises the chain the release actually ships
• ci-monorepo-test covers v9 on Python 3.12, and ci-monorepo-lint runs its
  circular-import check against v9
• Dependency updates: typer, ty, mypy, ruff, mkdocstrings-python,
  mkdocs-literate-nav, mkdocs-gen-files, griffe-inherited-docstrings