v1.4.0

[1.4.0] — 2026-06-10

Feature release: high bit depth and additional chroma samplings, plus the
fixes from a full audit. 8-bit 4:2:0 output with default parameters is
unchanged (all 198 pre-existing golden hashes reproduce bit-identically);
diMode=1 scene-change frames and diMode=2 edge columns intentionally
changed toward Avisynth parity (see Fixed).

Added

• 10/12/16-bit integer YUV input (u16 storage). IVTC decisions
  (field matching, decimation, scene logic) are bit-depth-deterministic:
  the decision kernels round at the 8-bit map level (pavgbScore), so a
  shifted 8-bit clip decides identically at any depth, while output
  pixels keep full precision. No external oracle exists for HBD
  (upstream is YV12-only); guarded by cross-bit-depth consistency tests.
• 4:2:2 and 4:4:4 chroma sampling across all kernels (4:1:1/4:4:0
  remain rejected).
• Real-content robustness + bit-depth-consistency harness
  (scripts/test_real_content.py).
• threshold parameter is now actually wired (upstream accepted it but
  hardcoded AdjPara(20); identical behaviour at the default 20).
• Golden coverage for the previously oracle-free axes: ref
  BOTTOM/ALL/NONE, diMode 0/1/2, blend 0/1 (new motion_flicker
  fixture — the first whose blocks actually pass the blend motion gate);
  compCn unit tests incl. the upstream uppercase-'N' quirk; the oracle
  parameter grid now lives in one module (scripts/param_grid.py).

Fixed

• 4:2:0 heights with height % 4 == 2 (e.g. 720×486) are now
  rejected. The field-interleaved chroma row mapping requires an even
  chroma-plane height; such clips previously caused out-of-bounds chroma
  reads on the default path, an out-of-bounds write in the blend path,
  and an uninitialized final chroma row (visible as cross-plane pixel
  contamination). Upstream has the identical latent bug, so nothing
  bit-compatible is lost.
• diMode=1 now always runs the full deinterlacer (Avisynth parity:
  only SIMPLE_BLUR/ONE_FIELD have the DrawPrevFrame scene-change
  shortcut). Previously a scene-change frame emitted a copy of the
  previous frame's matched fields instead.
• blend=1 is access-order independent: the blend kernel's reach into
  the previous block (base-1, weight 37/256) is now decided before
  rendering, so seeking produces the same pixels as linear playback.
• diMode=2 edge columns match upstream: the 3-tap motion test now
  replicates upstream's cross-row pointer wraps at x=0/x=w−1 instead of
  substituting 0.
• The 1.3.2 "Known issue" (diMode=1 SIMD pavgb vs scalar truncated
  average) was already fixed in the 4:2:2/4:4:4 work; the SIMD and
  scalar paths are verified identical across all kernels.
• Hardening: threshold/pthreshold validated to [0, 100000]
  (previously overflowed i32 internally in release builds); VFR clips
  and fps=24 on <2-frame clips rejected with clear errors; height capped
  at 8192; out-of-nominal-range 10/12-bit samples saturate instead of
  invoking UB; allocation-failure paths now set a proper filter error
  instead of returning garbage frames.

CI / packaging

• The integration suite (golden hashes, upstream comparison, determinism,
  bit depth) is now gating — it previously ran with a job-level
  continue-on-error and could never fail CI; a missing plugin now fails
  the run instead of silently skipping every test.
• Releases run the full test gate (Debug + ReleaseFast unit tests +
  integration suite against the ReleaseFast artifact) before anything is
  published; unit tests also run under ReleaseFast in CI.
• PyPI wheels now carry the full manylinux compatibility alias tags in
  WHEEL (the comment claimed it; the code wrote only the first alias).
• GitHub release bodies use the latest changelog section, not the whole
  file; pytest collection pinned to tests/integration.