The sensor does not care about iso, iso step is added after the readout.
Actually, that's not true -- and, presumably, at the heart of your misunderstanding.
Analog electronic (not digital) amplification / gain is applied to all ISO settings above the base ISO. It's like turning up the volume knob on your stereo. The readout is performed after that gain has been applied. And, just as your amplifier will start to produce more and more ugly-sounding distortion the louder and louder you crank that knob, your camera will produce more and more noise the higher you boost the ISO.
ISO 50 and highlight tone priority are two special cases. The exact same analog signal amplification (i.e., none) is applied with all three exposures: ISO 50, ISO 100, and ISO 200 w/ HTP. If you use the same shutter speed and aperture, you'll get the exact same raw file in all three cases. What changes is first the camera's metering system and second the digital (not analog) post-processing. That post-processing is quite simple, really...in the case of ISO 50, all the digital values from the initial sensor readout are exactly halved; with HTP, they're exactly doubled. (And, of course, with ISO 100, they're left as-is.) That's all simple integer math, too -- nothing fancy.
Hope that clarifies things somewhat....
Well, this was a long thread, and there's to little time...
I'll take this post from the start of the thread and start there. If I missed something in the in-between pages, shoot me!
I'm KIND of reading this as you're talking past each other, but anyway:
No, the ISO amplification is added after what is usually called the "read-out". The readout in itself is the analog coupling between the individual pixel charge storage where the charge is translated into a feed line voltage via the cell capacitance, a resistor and the reference level - one of the base definitions of a CMOS sensor.
The conversion from charge to voltage at pixel level is done at a fixed rate, that can only be adjusted by changing the pre-charge or the reference voltage feed, and you usually don't do that. Some machine vision sensors have facilities for this (absolute conversion rate calibration), but no sensors in the commercial area AFAIK. After this conversion has been done, the voltage is "transported" - shorted, connected to - a voltage amplifier that sits at the edge of the sensor, just before the signal leaves the sensor plate. The ISO amplifier.
In a Canon sensor, you usually have somewhere between 4-12 communication lines out from the sensor plate edge - each of those has it's own voltage amplifier where the ISO gain is applied. This is also what causes most of the vertical banding - the gain in those amplifiers aren't matched perfectly.
At base ISO, the "whites" are in the green channel usually quite close to the absolute clipping point of the individual pixel - the full well capacity. As you approach FWC, no more charges can be stored, no more voltage can be read out from the pixel. At this point, it doesn't matter if you try to lower the ISO amplifier gain, the data is still just a flatline, pinned at the maximum value of the pixel. Like trying to salvage a pure blown white in a jpg file - it's quite impossible.
So what would really be good to know if you're shooting raw is if the maximum pixel voltage that you can convert into an ADU value increases when shooting at one step below "normal" base ISO. And usually it does - by some 5%. The 5Dmk3 goes from about 67ke- to 70-71ke- The added range is mostly compression effects, distorted data - but that doesn't matter much in the highlights. On the other hand, 5% doesn't matter much anyway - +5% is the same as 0.07Ev or 1/14 Ev. Not much.
I have no idea how Canons HTP works in the transition from base ISO to higher ISO in newer cameras. If the camera is at ISO400, it's simple - just tell it to halve analog amplification (to ISO200) and compensate in post, pulling in the highlights as you do. In this case you don't lower the actual photometric exposure, you just adjust the amplification. The actual photoelectric conversion rate is constant.
But at base ISO this is meaningless, since there's no data to pull in. The pixels themselves are what they are, they can't send the amplifier any higher voltages, FWC sets the upper limit. Trying to lower amplification in that case is counterproductive, you're just hurting the shadow tone resolution.
compare it to an analog sound chain.
You have a microphone that can convert sound into a voltage, maximum output is 2v. This is the pixel.
You have an amplifier that can amplify the signal if needed (ISO gain, this is NOT in the pixel)
You have a loudspeaker that needs exactly 2v to play at the sound pressure that you want, no more and no less (your image brightness)
You adjust sound pressure at the output - image brightness - by adjusting the amplifier
Now if the microphone gives those maximum 2v from a 96dB sound pressure, screaming into the microphone at 102dB won't increase the maximum output. It will just distort the sound, by clipping. It doesn't matter if you turn the amplifier down, the distortion (clipping) is still there, since it's present at the source, the input of the amplifier. And there's nothing you can do about it except get another microphone with a lower base sensitivity (lower base ISO).
Actually it's quite easy to test - point a camera at something with graded highlight, take one shot at base ISO and one shot with exactly the same shutter speed and aperture at the faked ISO. Try to pull in the highlights in the raw file. Is there any difference?
If there's no difference, you could just as well have shot at base ISO with -1Ev compensation. You're just adjusting if you want to clip the highlight or the shadows, you gain absolutely nothing. It's a 1:1 trade-off.