fix: stop ignoring corpus/ so the refresh workflow can commit it

PLAN.md's design has corpus/ committed and chroma/+bm25/ regenerated
at CI time. The scaffold's .gitignore over-ignored corpus/, which
meant refresh.yml's `git add bundles.json corpus` silently dropped
the corpus and the changed-detection logic always reported "no
changes — skipping reindex and image build". Net result: refresh
would scrape successfully and then ship nothing.

chroma/ and bm25/ stay ignored — those are rebuilt by
`python -m rag.index --rebuild` before docker build copies them.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

.gitignore

@@ -2,11 +2,17 @@
 venv/
 .venv/
 
-# Regenerable from corpus + CI
+# Indexes are regenerated from corpus by `python -m rag.index --rebuild`
-corpus/
+# (run in CI before docker build). Don't commit them.
 chroma/
 bm25/
 
+# corpus/ IS committed — the weekly refresh workflow writes scraped
+# pages here and `git add bundles.json corpus`s them. The image-only
+# workflow then rebuilds indexes from the committed corpus without
+# re-scraping. Earlier the .gitignore silently ate `git add corpus`
+# and refresh.yml's commit step would always report "no changes".
+
 # Python detritus
 __pycache__/
 *.py[cod]