[kroq-gar78]

test_rendering_downloads_an_mp4 fails intermittently on CI with:

AssertionError: {'width': 320, 'height': 240, 'duration': 0.2, 'lit': 0}

indicating that the rendered frame is blank.

Likely cause

The MP4 is fine, but the check that plays it back is racy. The test loads the file into a <video> and draws it into a canvas on loadeddata. Chromium can fire loadeddata before the first decoded frame reaches the element, so drawImage draws nothing and every pixel reads as unlit.

Evidence, from playing one saved MP4 repeatedly with the old page under load (16 playbacks at once on 24 cores):

• lit == 0 in 10 of 400 runs. The other 390 drew the full frame from the same file.
• Each failure left the canvas fully transparent. A black frame from the encoder would be opaque, since the muxer paints black under every frame. So these failures drew nothing at all.
• It only occurs with several cores in use. Heavy load on a single core never triggered it, which fits a race between threads.
• 40 MP4s rendered by the test under the same load all decode with every frame lit in FFmpeg.

Fix

The playback page now waits for video.requestVideoFrameCallback, which fires once a frame has been presented, and then draws it once. If no frame arrives within 10 s, it returns an error. A frame normally arrives within 20 ms; the slowest run under load took 141 ms.

Under the same load the new page drew the full frame in 400 of 400 runs. Bad files still fail: an all-black video fails the lit check, and truncated or garbage files fail with a decode error.