adminandClaude Opus 5.5
2f13f9a0b9
In CI run 699 check 68k (real time behind a tab) got 21 of 120 frames. The beat ticked 100 times, but 83 frames were dropped as busy. Late in a long browser session the GPU behind a hidden tab took up to 1.5 s to finish a read back, so all three read-back buffers were still on their way when the next frame was due. Locally on melektaus-atx this showed only in the full run of modules 01 to 68, and not in every one. - The render loop no longer draws the screen while the page is hidden. Nobody sees it there, and the drawing was competing with the read backs for the slow GPU. - A recording of the editor copies the canvas, so it turns the drawing back on with `drawWhileHidden` while it runs. - `PictureReader.collect` reads the oldest buffer at once when the page is hidden and no buffer is free, instead of dropping the frame. The main thread waits for the GPU there, which nobody sees either. - More buffers are no fix: with 30 of them the frames piled up at the encoder instead (45 frames, 56 dropped). Tests: - render.test.ts: 18/18, the new test is red without the hidden check - Full run of modules 01 to 68 on melektaus-atx with a counting probe: read at once only 18 frames, 63 dropped; both parts 106 frames, 2 dropped - e2e 67 and 68 on melektaus-atx without probe: 68k green, 97 frames, 0 dropped. Red there only "a clip that starts later", the known timing wobble from task 697 - Gate on melektaus-atx: CHECK EXIT 0, 313 files, 4841 tests Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>