As with the previous version, which was using a hard-fork, the normalized transaction ID is computed only considering the non-malleable parts of a transaction, i.e., stripping the signatures before computing the hash of the transaction.
Is this proposal recursive?
Coinbase transaction
* n-txid = txid
Non-coinbase transactions
* replace sigScripts with empty strings
* replace txids in TxIns with n-txid for parents
The 2nd step is recursive starting from the coinbases.
In effect, the rule is that txids are what they would have been if n-txids had been used right from the start.