verify-real: bootstrap cdxt on demand for the fetch step

cdx_toolkit is no longer a standing dev dependency, so the real-path fetch check
installs it on demand (mirroring the HF fetch job's runtime install) instead of
assuming it in the venv.

Claude-Session: https://claude.ai/code/session_01EKnXafnPtBVLvtxUfU2GkR

scripts/verify_real.sh

@@ -22,6 +22,14 @@ CC_INDEX_MODE=cdn CC_CRAWLS="$CRAWL" CC_DOMAINS="$DOMAIN" CC_HOSTNAMES="" \
   CC_RANGES_OUT="$RANGES" $PY -m src.index_query
 
 N=$(($(wc -l < "$RANGES") - 1))
+# The fetch step runs the real `cdxt repackage` CLI (as the HF fetch job does). cdxt
+# is not a standing dev dependency — the estimate above uses the vendored SQL — so
+# install it on demand here if it is missing, mirroring the job's runtime install.
+if [ ! -x "$CDXT" ]; then
+  echo "== installing cdx_toolkit for the fetch step (not a standing dev dep) =="
+  uv pip install --python .venv \
+    "cdx_toolkit[hf] @ git+https://github.com/commoncrawl/cdx_toolkit.git@feat/warc-range-sources" uvloop
+fi
 echo "== fetch: $N records from the commoncrawl bucket via hf-cdn =="
 CDXT_UVLOOP=1 $CDXT -v repackage --target-source csv --csv-path "$RANGES" \
   --warc-download-prefix hf://buckets/commoncrawl/commoncrawl --hf-reader cdn \