the instagram carousel import on my taste engine broke, and the first clue was file sizes. the saved image came back a few kilobytes lighter than the one the cdn served, which smells like a re-encode. the real story was worse. the scraper regex targeted the old instagram polaris json, where width, height and url sat next to each other. the current schema changed. each slide now carries a display uri with separate dimensions, so the regex found zero matches in a huge window while the page clearly had eight slides.
then the refresh exposed the second bug. the cdn content-negotiates by user agent, so the downloader got an avif file, and the python step crashed before it could convert it. the image only came out right because an ffmpeg fallback decoded the avif into a jpeg. same correct file every time, absolutely deterministic. the fix worked, but through a path nobody intended.
the rebuild fixed the schema parsing, and the real blocker behind it. the service had a hardcoded localhost url for the headless browser while it ran on the container network, so the browser was unreachable until the config was corrected. after that the carousel test went from one row to eight real slides in about twenty seconds, with media sanity checks on the new files. the lesson is the boring one. when a scraper quietly returns nothing, check whether the page schema moved before you blame your code or your network.