The incremental gate was still counting failed re-reads. A failed publication or document read does not remove anything: the previous records are kept. Only a new publication or document that has no previous records can actually be lost by the run.
Count only those toward the failure rate, and keep excluding repositories already marked for a core retry. The 403/530 hosts on the last run are existing publications, so their re-read failures no longer refuse the dataset.
177 tests and prek pass.