Fix FFV1 writing: BGR0 instead of a format the encoder rejects

[kmatzen]

Fixes #70.

VideoWriter::open() forced FFV1 to AV_PIX_FMT_GBRP. The FFV1 encoder does not accept it:

[ffv1] Specified pixel format gbrp is not supported by the ffv1 encoder.
gbrp     -> FAIL: Invalid argument
yuv444p  -> OK
yuv420p  -> OK

(direct avcodec_open2 probe against libavcodec 62 — the encoder advertises gbrp9le and wider, but not plain 8-bit gbrp)

avcodec_open2 therefore always failed and FFV1 output never worked at all.

BGR0 keeps the stated intent — stay in RGB, no YUV conversion, no chroma subsampling — and matches the BGR24 cv::Mat input channel-for-channel.

How this stayed hidden

The two tests covering FFV1 called SKIP() when open() failed:

if (!writer.open(path, AV_CODEC_ID_FFV1, 320, 240, {30, 1})) {
    SKIP("FFV1 codec not available or pixel format not supported");
}

The message guesses at a missing codec. FFV1 is built into ffmpeg and was present the whole time; the format we handed it was wrong. Combined with ctest reporting "100% tests passed" regardless of skips (#66), a green build hid a completely dead code path.

Both are now REQUIRE. A present-but-unusable encoder is a bug, not an environment quirk.

Stronger assertion

The round-trip test allowed a tolerance of 2, with a comment attributing it to GBRP round-tripping through the scaler — a path that never actually ran. BGR0 never leaves RGB, so the round trip is bit-exact, and the test now checks equality:

CHECK(pixel[0] == 42);   // B
CHECK(pixel[1] == 84);   // G
CHECK(pixel[2] == 168);  // R

Passes.

Result

Suite goes from 37 passed / 2 skipped to 39 passed / 0 skipped — 1129 assertions, no skips.

I found this by investigating why those two tests skipped rather than working around them, while implementing the skip guard for #66. The skip check itself is a follow-up PR; it would have failed CI while this bug was live, which is the correct behaviour but needed this fixed first.

🤖 Generated with Claude Code