We have seen errors from operations using internal transaction api during stepdowns like

Attempted to run 'update' as a retryable write with session idbaa46b88-bdb2-4ca6-8b6a-7b60dce7e840 - 47DEQpj8HBSa+/TImW+5JCeuQeRkm5NMpJWZG3hSuFU= -  -  and transaction number 13 but the active transaction number on the session is 12 

From a discussion with Jack:

"Transactions can fail with certain errors like because of stepdowns, but they should be considered “transient errors” and the txn API should retry automatically on them. I think the problem is that particular error isn’t considered transient since we’re unstashing as a retryable write, which feels wrong."