idea for some easy improvements to auto exposure

Regarding point 1:
In that case it would still be useful with low-noise photos. Some random deviation in exposure would probably be preferrable to consistent overexposure. With 0.1% setting I didnt notice noise being a problem, so it's probably worthwile to push this further until problems with noise become obvious. I just wonder if one would even conciously notice such noise problems. Maybe they are already there and i'm just not noticing it?
I have some experience using this algorithm (ETTR + some fixed shift) for data visualisation. In that application I sometimes do exposure compensation for every pixel column individually, making the influence of noise an important factor because random column-wise exposure variations make the image have strong banding. I did indeed notice noise to be a problem in that setting. I found the noise problems could be alleviated by doing the following: instead of computing the exposure compensation value for a single pixel % value, compute it for a (narrow) range of % values, then take the average of them.
 
Regarding point 2:
Oh, great!

I thought more about your point regarding noise: As you said it such noise would be noticeable when switching between similar images. I don't recall having issues with this, so at least in my case I think the optimal pixel % would be below the current minimal allowed value.

I understood that.
I assume that you have either come to this conclusion because you tested this in the past or are going to test it, since i assume it requires changing only a single number. I just wanted to point out that my impression from practical use was that it would probable be helpful to allow setting lower values. Currently the minimum is 0.1%. If you think 0.001%  is too low, maybe 0.01 is fine? or maybe 0.03? It would probably depending on how noisy the images are and what kinds of subjects are being photographed. So even if testing sais that a certain number is the sensible limit, there are probably use cases that were not tested where lower numbers make sense.
 
I didn't mean Method 1 and 2 to be alternaties to each other. My intention is to use them together.
 
I don't know if what I wrote above about stabilising the result by averaging over a range of % values helps in this use case (photos). I dont know if you want to try that. I could try to code a simulation, but I dont have much time for such things right now.

Oh great! I'll make sure to check it out soon and report back.
 
Also, it just came to my mind that the app has already supported numbers below 0.1% since a long time. Namely, it allows setting it to 0%. It's just values between 0% and 0.1% that were not possible. By the logic that 0.001% is computing exposure based on 400 pixels (and thus subject to noise and other small image variations), 0% bases it on just a single pixel and is thus even more strongly affected by these issues.
 
In practice I traditionally kept bouncing between 0% and 0.1% depending on what types of images im culling, and almost never using anything higher. So my past usage pattern suggests that the problems that come with basing exposure on just a few pixels were a worthwile tradeoff for me in the past, even in the extreme case of just a single pixel, and that the optimal setting for me is somewhere in between.

Tested it, It's great! Many thanks!
I didnt run into any problems with it.
I decided to stick with 0.003 % for the time being.
The number of pixels shown as overexposed varied slightly (sometimes none at all) despite it exposing for a fixed proportion of overexposed pixels. But these variations were small, so they didnt cause me any problems in practice. I guess they could have been caused by limitations of the internal histogram accurracy mentioned above, or maybe the overexposure display uses a different definition of "clipped" then the exposure algorithm does. ("use camera limearity limit" for clipping definition was unchecked).

Is this a good place to report bugs of that feature?

1)
When I select "fixed exposure shift" then the "Manual adjust" field isnt greyed out


2)
When I set that manual adjust field to -2 and then look at an image that normally is computed to +1ev compensation then the autoexposure when triggered results in a program state where the exposure correction field reads "-1" but the image brightness corresponds to +1. If I manually set exposure correction to -1 then it works correctly. If I set manual adjust to -.5 then the autoexposure resulty in a correctly applied +.05 exposure correction. On another image that normally results in +0.03, using manual adjust of -.3 results in the field reading -.27 but the image brightness very close like a value of zero, but not exactly, probably like +.03.
So it seems like when the "manual adjust" field results in an overall negative exposure compensation value then it doesn't get applied to the image (image shows as if manual adjust was at zero) but it does get correctly shown in the exposure compensation field.
 
I think there would be two different sensible behaviours in such a case: either apply the manual adjustment fully, resulting in negative exposure compensation, or clip the compensation at zero (assuming there is no highlight reconstruction or other way to get data beyond the normal clipping point (e.g. a mode that uses per-channel clipping points).
I'm not sure which would provide better exposure consistency when there is a series of images where some would result in negative compensation and some wouldn't. Probably depends on whether there are images where sensor clipping point has clipped more then the % of pixels used to compute the auto-exposure compensation.
 

Nice!
I didn't encounter any additional bugs.