Tbf that's the actual service they offer. It's not like they are being sneaky.

To be sure, but anybody using CloudFlare needs to keep in the back of their head that everything going through CF is not only accessible to a hypothetical Bad Guy at CloudFlare it can also (and remember has, this is a real bug that happened) get exposed to unrelated parties on the Internet if CF mixes your data with somebody else's by mistake.

This makes CF seem fine for your different variants of popcorn.gif, your (subresource integrity checked) Javascript implementation of the VIC-20 computer, or a public blog post, and NOT so great for patient access to histology results, private web forums, banking, and many other things on the Web.

Of course they do that. How else would such a service possibly operate?

By offering two different models and letting you choose between them for each hostname under Cloudflare.

1) Injecting themselves on the application layer (the current model)

2) Injecting themselves on the network layer, passing through only encrypted traffic between the client and your servers

Not every feature they currently offer would still be possible in model #2, but most of them (including DDOS protection) can still work in a limited fashion.

Coming soon

That's very good news! The knowledge that a large part of my internet usage is MITM'ed by a single entity is continuously making me uncomfortable.