CI cannot run the Parquet integration tests

[gaurav]

.github/workflows/ci.yml's integration-test job runs pytest -v -m "integration and not slow"
but sets no Babel configuration, so it uses the public default. shared_downloader HEADs
duckdb/Concord.parquet, gets a 404, and skips 28 tests. The 13 that do run only exercise the
NodeNorm API.

The job is green and covers none of the DuckDB query layer, the downloader's real network paths,
or the cache-refresh logic.

Options

• Set BABEL_RELEASES_URL (and BABEL_VERSION) from a repository secret pointing at a tree that
  publishes the files. Note the runner would pull ~626 MB for Concord, so keep excluding slow.
• Or close this as obsolete once Publish the DuckDB Parquet files in public Babel releases #16 lands and the public default works — that is the better
  outcome, since it needs no secret at all.

Blocked on, and possibly resolved by, #16.

[gaurav]

Sizing correction from an end-to-end run against 2026jul22: Concord.parquet is 4.6 GB, not the ~626 MB an older fixture docstring claimed (fixed in 37e9fac). Its tests are not marked slow, so pointing this job at a Babel that publishes the Parquet files would have the runner pull 4.6 GB on every push to main and every weekly cron — plus Identifiers.parquet again if slow is ever enabled.

Worth deciding up front whether that belongs on a hosted runner at all, or whether these tests want a cached/self-hosted runner, a cut-down Parquet fixture, or a schedule less frequent than every push.