Add a source-request loop for what agents can't fetch

Datawrapper's blog 403s, Medium is unreachable, JSTOR wants a session,
and a search snippet is a paraphrase rather than a source's words. The
failure mode is guessing, or silently writing a weaker page.

docs/source-requests.md now holds a fetch queue with the specific
question each link needs to answer, plus a register of claims already
live on the site that rest on recall rather than a page anyone opened —
starting with the riskiest, an inference about which mistakes the
Economist piece owns up to. docs/sources/ is the drop zone.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01ULE5RRxdQE1ebwefEd2eCM

docs/sources/README.md

@@ -0,0 +1,19 @@
+# Fetched sources
+
+Drop zone for pages an agent couldn't reach itself. One file per source,
+named for the row it answers — e.g. `datawrapper-pie-charts.md`,
+`economist-mistakes.md`.
+
+No format requirements. Paste the relevant passages, or just write what the
+page says in a couple of sentences. Either works; the point is that a claim
+stops resting on recall.
+
+Worth stating at the top of each file: **where it came from and when**, so a
+later reader can tell a 2018 position from a 2026 revision. Datawrapper's
+dual-axis post is exactly this trap — it was rewritten years after the version
+everyone quotes.
+
+Keep files to the passages that matter. This repo is public, and these notes
+exist to check claims rather than to reproduce anyone's writing.
+
+See `../source-requests.md` for what's wanted and why.