Task generation fails when the Evergreen API returns a transient 401 or 404. Neither is in `evergreen.py`'s default retry list, so the error propagates and kills the task. These do happen spuriously — we have seen a 401 on the second page of a paginated `tasks_by_build` call whose first page succeeded on the same session with the same credentials.

Retry them, but on a smaller budget than the 5xx codes. Unlike those, a 401 or 404 is usually genuine, so spending the full ~21 minute budget before reporting it is wasteful. `TransientAuthRetry` allows 3 attempts spaced a flat 120s apart, a ~6 minute ceiling; 5xx and 429 keep the full budget.

Behavior on exhaustion is unchanged: callers still get the same `HTTPError`, just later.

Tests drive real retry chains rather than hand-built state, and cover the cap, the spacing, and that the 5xx budget is unaffected.

One limitation worth knowing: the 3 attempts are a budget for the whole call, not per page, so a paginated fetch that blips on several different pages shares them.