the failed-repo retry pass routed every getRepo through the relay's 302, concentrating the whole pass on bsky.network from the firehose's own IP (suspect in the 08-14 4h ingest stall). upstream's retryRunner.download dials the stamped PDS first and uses the relay only when the stamp is unroutable or the stamped host authoritatively lacks the repo; fallback success restamps the route. ported, with upstream's foreign-embedded-DID guard for the direct path, two counters, and regression tests that pin the stale-stamp 404 -> relay fallback -> restamp path (without which a migrated repo would be marked terminal-complete undownloaded). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>