The original implementation also involved sending a "safety voucher" with each photo uploaded to iCloud, which contained a thumbnail of the photo as well as some other metadata.

The vouchers were encrypted, and could only be decrypted if there were, I believe, 30 independent matches against their CSAM hash table in the cloud. At that point the vouchers could be decrypted and reviewed by a human as a check against false-positives.

It sounds like with a raw byte hash they might be able to match a photo against a list of CSAM hashes, but they wouldn't be able to do the human review of the photo's contents because of E2E.