Cache audio waveforms and rescue oversized GIF previews

docs/media-scraper.md

@@ -130,7 +130,12 @@ Previews run after every walk:
   10 s timeout, one attempt).
 - Animated GIF previews are stored as a static first frame (`ffmpeg`, 480px WebP); the original
   animated GIF URL stays in `preview_url` for hover playback. Tenor embeds without a thumbnail are
-  resolved through Tenor's oEmbed.
+  resolved through Tenor's oEmbed. An animated GIF over the size cap is framed by ffmpeg reading
+  the stream directly, so it is not lost to `too large`.
+- Audio attachments and voice messages get a waveform strip
+  (`ffmpeg -filter_complex showwavespic`) as their preview, rendered by ffmpeg from the URL.
+- Transient network failures are left pending for the next backfill pass; only skips (4xx, too
+  large, no source) are marked permanent.
 - One attempt per row, no retries; a failure marks `preview_error` and never fails the scrape.
 - Stored by hash in `thumbs/<sha256>.<ext>`, where the extension follows the response type
   (`preview_ext`, default webp); `preview_hash` and `preview_bytes` land on the row, and the

docs/media.md

@@ -91,8 +91,9 @@ media-scrape/db/
 - Third-party embeds are **never downloaded as media**: YouTube, Instagram, X, TikTok, and every
   other embed stays a link. Its thumbnail is cached (below); the video itself is never fetched.
 - Previews are cached in `thumbs/` after every walk, in both modes: Discord-hosted images, GIFs,
-  and video files (the proxy renders a frame, resized to 320px WebP) plus third-party embed
-  thumbnails (YouTube, X, TikTok, and so on). The
+  and video files (the proxy renders a frame, resized to 320px WebP), third-party embed thumbnails
+  (YouTube, X, TikTok, and so on), and audio, which gets a waveform strip so audio tiles are not
+  blank. The
   stored bytes keep their real type (`preview_ext`), so JPEG provider thumbnails are served as
   `image/jpeg`, not mislabeled WebP. Embeds without a thumbnail in Discord's data are resolved
   through the provider's oEmbed/fxtwitter API and finally the page's `og:image`. Previews count