Empty accounts (defined as having no nonce, no balance, no code and no storage) exist as a consequence of design flaws in early versions of the EVM. After an unknown attacker created large numbers of them during the Shanghai DOS attacks, EIP-161 both banned the creation of new empty accounts and provided that the old ones would be deleted if they were “touched”.
Vitalik Buterin then bulk touched all the accounts created by the Shanghai DOS attacker, but neglected to deal with 196 that had been created by accident.
These remaining empty accounts were largely forgotten about, but their continued existance raised the threat of consensus issues with this obscure corner of the protocol. This is best illustrated by this comment in the Geth source code, which incorrectly stated that
STATICCALLs and empty accounts can never interact on Mainnet. Additionally there were edge cases involving empty accounts with storage, that I have been unable to find a test for.
I deleted every single remaining empty account in block 14,049,881. This makes all the edge cases that never happened impossible, retroactively vindicates the Geth comment.
However, the Ethereum specification and testsuite still requires implementors to implement EIP-161 state clearing with all the impossible edgecases even where it is irrelevant. As an extreme, a post-merge only client would have to implement state-clearing solely to pass the Ethereum testsuite. The EIP removes as much of this technical debt as possible.
The EIP requires that certain facts about Mainnet be validated in order to be equivalent to EIP-161. 1 and 2 can be validated by dumping the state at the start of Byzantium to get the list of all 196 accounts that existed post-Byzantium and checking that they were deleted in the transactions named in the EIP.
3 is slightly more complicated. EIP-161 is usually implemented by marking an account as “touched” and then checking if it is still empty at the end of the transaction. Condition 3 can be checked by testing that the “if still empty” never fails during a sync to the start of Byzantium.
I have performed these checks by informally modifying the akula client, but something more rigorous will be needed before the EIP becomes final.