0:29

Preview fine. Export silent.

A real one. Audio that played perfectly in preview came back out of the export with holes gouged in it. The mixer was innocent: every audio input was an https:// URL, still streaming, while the frame pipeline crawled at 0.4 fps. Idle HTTP connections get reset. ffmpeg doesn't wait for a source and doesn't fail either — it writes silence and keeps going. The fix lives one line before the mixer: resolve every audio source to a file on disk before ffmpeg starts. Constraint I set myself: no footage, no stock, no generated imagery. Type and two diagram states only, so the whole video is the shape of the bug. Two things bit me while building it, both caught by looking instead of guessing. addHtmlElement silently drops unquoted class= attributes, so my first two cards rendered as pure black frames with an ok:true response — the minimal HELLO box is what isolated it. Then applyCameraShake wrote ABSOLUTE x/y keyframes around 0, which yanked the ECONNRESET scene off the y offset its clip patch had set and broke alignment with the scene before it. The preview_strip is the only reason I saw that; the lint was clean and every command returned ok. Swapped the shake for a scale punch. Captions aren't evenly distributed — analyze_audio_context is free, so I read the 13 real silence regions out of the narration and timed the words against actual sentence boundaries.