The hive-glob view must LIST every object under era5/daily/ at view bind —
~1k objects per tile, which crossed the request timeout once the lake hit
390 tiles (seen live: prod /query timing out). Bucket mode now resolves the
newest Iceberg metadata JSON (one single-page LIST) and iceberg_scans it, so
the file list comes from table metadata; local mode (tests, small trees)
keeps the plain glob. The iceberg extension is baked next to httpfs, as the
runtime user.
