wait() (sync + async) only stopped on finished/stale/committed, so a job the
server already reported as state=failed was polled until the (default 3600s)
timeout, then raised a misleading TimeoutError instead of the real cause. A
doomed backfill -- e.g. a multi-column REFRESH COLUMN of a scalar UDF -- hung
the client even though get_job surfaced the failure within ~3s.
Add a terminal failed branch that raises JobFailedError carrying the server
error, exported from the package. Verified end-to-end against the cluster:
raises in 3.6s instead of hanging. Unit-tested with a mock conn (sync+async,
failure + success + committed paths).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
3a4cdb7aff