The whole point of the CSAM stuff was that it would allow for end to end encryption while not turning Apple’s ecosystem into preferred tool of child pornographers.
Apple poorly communicated the feature, then the EFF put out a deliberately misguided written hitpiece that conflated parental controls with CSAM, and started an online freak out. The “privacy activists” won, and your data is sitting on Apple servers with Apple’s managed encryption keys outside of your control today.
In their proposal they would scan your files on device, which is fundamentally different. Initially they would not run the scanning when icloud upload was disabled but how long would that last for?
Phrases like “it would not be doing any scanning on-device” don’t have any precise meaning. Scanning is a series of operations including hashing and cryptographic calculations plus a voucher upload and decryption. All of the former operations are happening on the device, only the latter happens on the server. So in fact a significant fraction of the scanning is indeed happening on your device. And this two-computer design isn’t being used to preserve your privacy: it’s designed this way solely to prevent you (the device owner) from knowing whether your files match the database. Without that requirement, the system would be much simpler: it would download a hash database and simply send a notification to iCloud whenever (a sufficient number of) local files hash to values matching the database.
Also at the point that the image hash matches, Apple thinks it is CSAM (and it probably is). Doing it locally lets them avoid storing it, which they definitely do not want to do.