As far as I'm aware, HTP has always worked for raw files. Are you sure you're not confusing it with auto lighting optimizer?
It's essentially shooting 1 ISO stop lower then applying a selective curve to boost the shadows/midtones, thus giving you an extra stop of highlight wiggle room - or at least this is how I've always understood it. The data is 'baked' into the raw file, but you'd be just as well off underexposing and applying your own curve adjustment in post. The result is increased shadow noise.
Someone correct me if I'm wrong.
The HTP processing
is not 'baked into the RAW file'. You're correct about what it does - reducing the ISO by one stop, then applying a tone curve to selectively push (meaning 'highlight preservation' at the cost of more shadow noise). The only thing baked into the RAW file is the one-stop of underexposure (which, I presume, is what Chuck Westfall meant when he stated that HTP affects the RAW image). In the metadata, the selected ISO is recorded (which is different than the ISO actually used), and the HTP flag is set.
When you open in the image in most RAW converters, the converter sees the flag and applies a tone curve to boost the exposure. Canon's DPP applies the same tone curve used in-camera. LR (ACR) uses Adobe's interpretation of that tone curve. The fact that you see a difference when you push the resulting images in your test is no surprise - you're comparing an exposure change in post added
to Adobe's interpretation of the selective tone curve when HTP was used.
If you open an HTP-shot image in a RAW converter that doesn't support Canon's HTP flag, you'll just see a darker image - that same one-stop of underexposure that's the only effect on the RAW image data itself. That one-stop reduction does mean better preservation of the highlights, but you caould achieve the same preservation by applying a -1 EC before the shot. It's just that HTP may appear to do a better job, because it's using a complex tone curve that's difficult to manually create.