[ThomasKirst]

Hi, and thanks for cinetic! We used 1.0.0 to make a 29 s product film for our app: a 16:9 master plus 9:16 and 1:1 re-layouts. The probe check caught a colour problem on the way, and this PR fixes it. It is small and self-contained. A second PR with two smaller things (#4) and an issue with notes from the project (#5) are separate.

The film, the debugging and these texts were made together with Claude (Opus 5.5) in Claude Code. Each measurement below comes from a small ffmpeg test that is easy to rerun; happy to share the commands.

Blurred masters: shifted colours and missing tags (accumulate.py)

With ffmpeg 9.0.2 (the current Homebrew version), every render.sh --blur master failed probe.py's colour check with bt709/unknown/unknown/tv.

The cause is the encoder call:

• accumulate.py pipes finished BT.709 planes as -f rawvideo -i - and gives the colour tags only as output options.
• ffmpeg 9 then reads the untagged pipe as BT.601 and auto-inserts a conversion to the requested matrix (auto_scale_0 in -v verbose).
• So the colours shift, and primaries and transfer still come out as unknown.

The fix moves the four tags before -i, so they describe the planes as they are.

64×64 flat #00C2E0, piped as BT.709 yuv420p, crf 0	Y / U / V decoded	tags
piped planes	149 / 161 / 42	–
before (tags on the output)	162 / 151 / 42	bt709/unknown/unknown/tv
after (tags on the input)	149 / 161 / 42, bit-exact	bt709/bt709/bt709/tv

On accumulate.py itself (8 testsrc2 PNGs, --uniform 2) the tags go from bt709/unknown/unknown to bt709/bt709/bt709.

Sharp hf-finish.sh encodes: missing tags (refs #2)

This is point 3 of #2, and I could reproduce it on ffmpeg 9.0.2 as well: the PNG → H.264 encode comes out bt709/unknown/unknown/tv, because the filter graph's frame properties override the output-side tags. Following the suggestion in #2, the -vf chain now ends in setparams=color_primaries=bt709:color_trc=bt709:colorspace=bt709:range=tv. The decoded pixels are identical before and after (same MD5); only the tags change.

Points 1 and 2 of #2 (bash 3.2, BSD wc) are not part of this PR. For reference: bash -n scripts/hf-finish.sh already fails under macOS's /bin/bash 3.2 on main.

Docs

• references/finishing.md §4: the corrected recipe, with a short note on both cases (filter graph vs. raw pipe).
• probe.py: the hint now points to that recipe.
• CHANGELOG.md: two entries under [Unreleased].

Checks

• lint-film.test.mjs passes.
• pick.test.py passes.
• py_compile passes for all scripts.
• In our film, the master and both variants pass probe.py with these changes.