Ingest a source into the wiki — a durable summary, a raw backstop, folded into every page it touches.

Resolve the vault per references/vault.md, then read its AGENTS.md and run this in the vault's own terms.

Get the raw

Start from whichever entry point you were invoked with:

A source given — a URL in the argument, or the source under discussion. Fetch it and save the full text verbatim to the raw inbox (Markdown, PDF, or image). Interactive by default: read it, talk the takeaways over with the user, then file.
- A video URL — YouTube, Vimeo, TED, or any other site yt-dlp supports: its full text is the transcript. Run scripts/fetch_transcript.sh <url>; its header documents the output. Exit 4: rerun with the title's language as the second argument, or ask the user when the title doesn't show it. Any other failure: report the script's error output and stop — a video summary comes from its transcript alone.
No argument — sweep the raw inbox for clips the user captured but never filed (Web Clipper output, no processed stamp). Batch them with less back-and-forth.

File it

For each raw item:

Write a self-contained summary into the summary layer — enough that it survives link rot.
Fold the source into every evergreen page it touches — the concepts it informs and the entities it names, often many. Update every one that exists; link the rest as candidates.
Stamp the raw item processed so the next sweep skips it. Never edit the raw beyond the stamp.