Decide what a pipeline triage agent may do

A nightly data pipeline fails a few times a month, and someone has to decide whether to rerun it, backfill it, and who to warn. An agent makes the first call on each failure, and it may never start a backfill on its own.

Write triage_pipeline_failure(failure_class, is_idempotent, prior_retries, downstream_consumers): the rule your agent follows for each failure.

• failure_class (str): the diagnosed cause.
• is_idempotent (bool): the failed task is safe to run twice, because it overwrites rather than appends.
• prior_retries (int): how many times the task was already retried.
• downstream_consumers (int): how many downstream consumers read the stale tables.

It returns one of the values in the rules below. Apply them in order. The first one that matches decides:

1. Trim and lowercase failure_class. If it is not infra, upstream_late, schema_drift or code: "escalate_to_human".
2. The class is infra or upstream_late, the task is idempotent, and prior_retries is 0 or less: "retry_once".
3. downstream_consumers is above 0: "notify_consumers_and_await_approval".
4. Otherwise: "await_backfill_approval".