Meter the lowest histogram bin in auto exposure

[stuartparmenter]

Objective

Auto exposure can make a sudden adjustment when most of the view becomes dark.

To work out how much to brighten or darken the image, auto exposure groups pixels by luminance in a histogram, filters out a percentage of the darkest and brightest samples, and averages the rest.

The problem is that pixels below AutoExposure::range count toward the filtering percentages but get skipped when calculating the average. As the view gets darker, the average still depends on the brighter samples. Once those are all filtered out, it falls back to the minimum luminance, making the target exposure jump.

Solution

Include the darkest pixels in the average too. They go into histogram bin 0 and are treated as having the minimum luminance. That lets the average decrease as more of the view gets dark, without the jump when only dark samples remain.

The filter and metering_mask settings can still be used to leave dark pixels out.

Testing

In addition to testing as part of my Zorah work, I built tests that add more dark pixels in steps and check the exposure against expected values. They reproduce the jump before the fix and show gradual changes afterward. Also checked different luminance ranges, filter settings, and metering masks.

---

This was AI assisted w/ Fable 5.1 & Astra as part of my work to get the Zorah scene working, which has a lot of very bright and very dark areas, and I was noticing weird behavior after I added auto exposure in, so had the LLM help track it down, put together a fix, and then visually confirm it was fixed. Then built explicit tests to validate exact measurable behavior. Reviewed everything carefully & cleaned it up a bit before posting.

[alice-i-cecile]

Can you leave a quick migration guide note? This is the sort of thing that people are going to be baffled by and struggle to track down, even though I think that this is a reasonable refinement.

Otherwise I like this PR and think it's correct: I have some exposure to this stuff via photography and think that this is a better approach than our previous one.