Well, I’ve learned a few things about how Boulder’s nonce generation is implemented 
Apologies for leading you down the wrong path initially.
The short answer here is that the nonces you submitted that generated a bad nonce error were expired.
Like @jsha mentions in the thread you linked the nonces do not have a fixed validity period (e.g. good for 1d, 4h, etc). Rather Boulder has a fixed number of nonces that it keeps in-play, and if more than that number are requested some of the existing nonces fall out of validity and are in-effect expired.
Based on current usage volume I’d estimate we start rolling nonces out of validity inside of ~1-2 hours for production. Since your requests were ~4 hours apart the nonces had expired. I think the ultimate take-away should be to view the bad nonce errors as non-fatal and have your systems recover by refreshing the nonce used.