Per-process gating won’t help if the quota is actually scoped to the API key, service account, or consumer project: several processes or customer credentials under the same project can aggregate above 800 reads/min even when each process stays at 720. I’d group your request metrics by project/credential and test one credential with retries disabled. The 429 after 60s idle could also be delayed rolling-window accounting or a retry from another worker, so confirm the effective scope before assuming a lower hidden quota.