test(tools): cover frames with an odd width or height

[lusoris]

Summary

This adds the regression test the fork was missing for the defect of our upstream Netflix/vmaf PR #1604. No reader code changes.

A subsampled plane occupies ceil(width / dec) * ceil(height / dec) samples in a raw or y4m file. Upstream's direct readers read the floor, so part of every odd-sized frame stays in the stream, the next frame is read from the wrong offset, and the tool crashes. The fork is not affected: VmafPicture carries ceiling chroma (vmaf_chroma_extent(), Research-0094), and the buffered reader the vmaf tool uses reads a frame in one piece. Nothing tested that, so a rebase that restored floor chroma would have gone unnoticed here.

test_video_input_odd_dims derives every sample from its plane, row, column and frame, reads three frames and then requires the end of the clip, through both reader entry points (video_input_fetch_into_vmaf_picture() and video_input_fetch_frame()):

• 20x20 4:2:0 as the control, raw and y4m;
• 19x19 4:2:0, raw and y4m (the case of chore(deps): Update dependency tox to v4.64.4 #1604);
• 19x20 and 20x19 4:2:0, and 19x19 4:2:2 raw;
• a clip cut short inside its last frame, which must return -1, not 0 and not 1.

Type

• feat — new feature
• fix — bug fix
• perf — performance improvement
• refactor — no behavior change
• docs — documentation only
• test — test-only
• build / ci — tooling / infra
• port — cherry-pick from upstream Netflix/vmaf
• sycl / cuda / simd — backend-specific

Checklist

• Commits follow Conventional Commits (the commit-msg hook enforces this).
• make format && make lint is green locally. Not run in full. Run instead: clang-format --dry-run -Werror on the new file (clean), scripts/ci/tidy-ratchet.py --lane cpu --only core/test/test_video_input_odd_dims.c (0 warnings), and the pre-commit hook set (passed).
• Unit tests pass: fast suite on a CPU build, 191 passed, 0 failed, 2 skipped, and on a CUDA build (gcc, nvcc, RTX 4090), 249 passed, 0 failed, 1 skipped; the new test also passes under ASan + UBSan + LeakSanitizer.
• If I touched any SIMD/GPU code path, I ran /cross-backend-diff and the worst ULP is ≤ 2. None touched.
• If I touched a feature extractor with SIMD/GPU twins, I either updated every twin or listed the gap under "Known follow-ups" below. None touched.
• If I added a new .c / .cpp / .cu / .h / .hpp, it has the appropriate license header (see CONTRIBUTING.md).
• If this is a breaking change, the commit message uses ! or BREAKING CHANGE: and the migration path is documented below.
• If this PR adds an ADR, the ADR row lives in docs/adr/_index_fragments/<NNNN-slug>.md and the slug is appended to docs/adr/_index_fragments/_order.txt.

Bug-status hygiene (ADR-0165)

• docs/state.md updated: T-VIDINPUT-ODD-DIMENSION-READBACK-UNTESTED-2026-10-01 under Recently closed.

Netflix golden-data gate (ADR-0024)

• I did not modify any assertAlmostEqual(...) score in the Netflix golden Python tests.
• If I believe a golden value must change, I have explained why below AND pinged @lusoris for a CODEOWNERS exception.

Deep-dive deliverables (ADR-0108)

• Research digest — no digest needed: test-only change.
• Decision matrix — no alternatives: only-one-way fix already in tree; upstream's test could not be taken as it is, because it expects floor chroma in the picture.
• AGENTS.md invariant note — no rebase-sensitive invariants beyond the rebase note below.
• Reproducer / smoke-test command — pasted below under "Reproducer".
• CHANGELOG fragment — changelog.d/changed/test-video-input-odd-dimensions.md.
• Rebase note — entry added to docs/rebase-notes.md; the older note that offered upstream's test file is corrected.

Reproducer

CC=gcc meson setup build-san core -Db_lto=false -Denable_cuda=false -Denable_sycl=false \
  -Denable_float=true -Db_sanitize=address,undefined -Dbuildtype=debugoptimized
ninja -C build-san test/test_video_input_odd_dims
ASAN_OPTIONS=detect_leaks=1:halt_on_error=1 UBSAN_OPTIONS=halt_on_error=1 \
  build-san/test/test_video_input_odd_dims
# 4 tests run, 4 passed

Negative controls, run locally and not committed:

• floor chroma in vmaf_chroma_extent() (upstream's picture geometry): test_odd_420_reads_back fails with the picture does not carry the planes the file stores;
• floor chroma in yuv_input_set_plane_geometry(): AddressSanitizer stops the test in the buffered reader.

End to end on master c7f28317f: three 19x19 4:2:0 frames in a y4m container give psnr_y 32.360917 / 34.161521 / 33.848468, the values upstream reports with its fix.

Known follow-ups

• Observed, not changed: the tool's odd-dimension refusal (validate_chroma_alignment() in core/tools/vmaf.cpp) tests frame_w / frame_h. The y4m reader pads those to a multiple of 16, so the refusal fires for raw input only (odd width 19 not allowed for chroma-subsampled format), and odd-sized y4m input is read and scored. Both readers are frame-aligned, so no score is wrong; whether odd y4m input should be refused as well is a maintainer decision. The docs/state.md row records it.
• The y4m 4:2:2 case is left out on purpose: the y4m reader resamples C422 chroma to the jpeg siting, so its output is not the file's samples.