Testing

python -m pytest -q runs everything; nothing is skipped. The integration tests need FFmpeg on PATH (to synthesise fixtures and for ffmpeg-skill) and an ffmpeg-skill checkout (COLOR_GRADING_FFMPEG_SKILL_DIR, or ../ffmpeg-skill, ./vendor/ffmpeg-skill, ~/.claude/skills/ffmpeg-skill); a missing one fails the session (not a skip).

file	covers
tests/test_unit.py	request schema (every operation's parameters, arity, references, forbidden fields), graph (order, cycles, unreachable, LUT content-hash identity override, determinism), canonical JSON, error table, file-name rules, path policy (input / LUT roots independent), symlink escapes, Windows names
tests/test_security.py	no shell in source, script allow-list (probe.py/color.py only), argv formatting (fmt_num), parameter injection through every operation's parameters, unsafe output paths / roots / symlinks through the CLI, LUT path policy separate from input path policy, input overwrite, existing output, malformed JSON, command/argv fields anywhere in the document, ffmpeg-skill directory handling, minimal environment, argv builder audit (every value is a fixed flag / formatted number / validated enum / absolute path)
tests/test_contract.py	contract ⇔ implementation, skill = contract, doctor statuses (supported / unsupported / unknown), doctor without ffmpeg-skill, capability table
tests/test_integration.py	per operation positive + negative (HDR_TO_SDR incl. non-HDR source refusal and force, LUT_APPLY incl. wrong extension / oversized / outside allowed roots, RETAG for all four targets, STRIP_DOVI on HEVC and its refusal on H.264, PRIMARY_CORRECTION incl. defaults-are-near-identity, deterministic desaturation, out-of-range rejection and observed measurements, plus gamma/lift/gain/levels_*/curves measurable effects, invalid curves choice, and levels_in_black/levels_out_black ordering rejection), chained pipelines HDR_TO_SDR → LUT_APPLY and PRIMARY_CORRECTION → LUT_APPLY, dry run writes nothing, re-run reuse + --no-reuse + tampered intermediate, invalid inputs (no video stream, corrupt file), output container mismatch, missing output, expectation failure removes the output, timeout, SIGINT cancellation, CLI validate / exit codes / human output, deterministic pixel-level checks for LUT_APPLY and PRIMARY_CORRECTION

Fixtures (tests/fixtures/generate.py) are synthesised with ffmpeg at test time: a 3 s 160×90 SDR H.264+AAC clip (solid colour), a 2 s SDR H.264 clip with no audio, a 2 s SDR HEVC clip (no Dolby Vision side data), a 2 s HEVC Main10 clip tagged BT.2020/PQ (HDR10), a 2 s mono PCM WAV (no video stream), a text file, and a size-2 "invert" .cube LUT (a genuinely affine map: trilinear/tetrahedral interpolation of its 8 corners reproduces output = 1 - input exactly at every point in the cube, so the resulting colour shift is deterministic and numerically checkable — never a subjective "looks graded" judgement). Expected values in the tests follow from this construction.

What real-media verification does and does not prove

• HDR_TO_SDR: run against the HEVC10/PQ/BT.2020 fixture; the output is probed and must report hdr: false (ffmpeg-skill/probe's own HDR detection, not this skill's guess) — a real, deterministic check that tone mapping actually ran, not merely that ffmpeg exited 0.
• LUT_APPLY: run with the invert LUT against the solid-colour fixture; the output's average colour (sampled by scaling the frame to 1×1, an area/box-filter average) must equal 255 - average colour of the input within a documented tolerance (chroma subsampling and lossy H.264 encoding introduce a few counts of error). This is the "pixel statistics" check STEP 19 of the design brief asks for, applied to a mathematically known transform — never a subjective aesthetic judgement.
• RETAG: run for all four targets against the SDR fixture; the output's (color_space, color_primaries, color_transfer) triple, as reported by ffmpeg-skill/probe, must equal ffmpeg-skill's own documented mapping (model.RETAG_TAGS) exactly.
• STRIP_DOVI: run against the plain HEVC fixture; the operation succeeds (HEVC is accepted) and the output reports no Dolby Vision side data. Limitation, stated plainly: this repository cannot synthesise a genuine Dolby Vision RPU (profile 8.4) with plain ffmpeg/lavfi sources, so the test does not prove that an existing RPU is stripped — only that the operation runs end-to-end on an HEVC input and the output is validated the same way a real Dolby Vision source's output would be. STRIP_DOVI on a non-HEVC (H.264) source is tested as the negative case ffmpeg-skill itself enforces (TOOL_ERROR).
• PRIMARY_CORRECTION: run against the solid-colour fixture. All-default parameters (an intended no-op) leaves the sampled average colour within a documented tolerance of the input (chained-filter rounding plus the x264 re-encode itself cost a few levels even at identity, which is real, measured drift, not a colour shift); saturation=0.0 is checked as a deterministic, measurable transform (near-equal sampled R/G/B — grayscale is a mathematical fact about the output, not a "look"); an out-of-range parameter (e.g. exposure=10.0) is rejected with INVALID_REQUEST before any tool runs. ffmpeg-skill's own before/after measurements (signalstats-based, real numbers) are asserted present on both results[].measurements and the matching provenance chain entry, and are asserted to differ between input and output for a parameter combination that measurably changes the image — this is the "observed, not judged" requirement verified end to end, not merely unit-tested against a mocked tool response.

Static checks used in development: python -m pyflakes src tests, python -m compileall src tests.