OpenImageIO: [BUG] OIIO seems to consider that the alpha channel is gamma-compressed when reading/writing 8-bit formats

Describe the bug

Use the attached images ramp.png and solidblack.png. ramp.png has a linear alpha variation from left to right (8-bit values checked with the GIMP) and do:

oiiotool ramp.png --ch "R,G,B" -o rampoiiorgb.png # result way too dark
oiiotool --no-autopremult ramp.png --tocolorspace linear --premult --ch "R,G,B" --tocolorspace sRGB -o rampoiioopaque.png # GOOD expected result
oiiotool ramp.png --ch "R,G,B" -o rampoiiorgb.exr # result not fully OK, the red and black bands are too narrow
oiiotool rampoiiorgb.exr --tocolorspace sRGB -o rampoiiorgbviaexr.png # same

# composite over solid black
oiiotool ramp.png solidblack.png -over -o rampoversolidblackoiio.png # result too dark
oiiotool --no-autopremult ramp.png --tocolorspace linear --premult solidblack.png -over --tocolorspace sRGB -o rampoversolisblackoiio2.png # GOOF expected result
oiiotool ramp.png solidblack.png -over -o rampoversolidblackoiio.exr # result not fully OK
oiiotool rampoversolidblackoiio.exr --tocolorspace sRGB -o rampoversolidblackoiioviaexr.png # result OK

I checked manually the values at x=960, y=540:

• ramp.png has RGBA value (177,0,196,128) (50% alpha, unassociated because PNG)
• rampoiiorgb.png and rampoversolidblackoiio.png have RGB value (89,0,98)
• rampopaque.png (attached image) has RGB value (129,0,143) (it was obtained by linearizing RGB from ramp.png, then compositing over black with the linear alpha or equivalently multiply by alpha, then delinearizing RGB)
• rampoiioopaque.png and rampoversolisblackoiio2.png have RGB value (129,0,144), close to what I compute by hand
• rampoiiorgbviaexr.png and rampoversolidblackoiioviaexr.png have RGB value (159,0,167), which I cannot explain, where do these values come from?
• R_good = ((177/255)^2.2 * (128/255))^(1/2.2) * 255 = 129 <-- I should get this value
• B_good = ((196/255)^2.2 * (128/255))^(1/2.2) * 255 = 143 <-- id
• R_bad = ((177/255)^2.2 * (128/255)^2.2)^(1/2.2) * 255 = 89 <-- I get that instead: alpha was not considered linear it seems! Or alpha was applied on the nonlinear RGB: this is also simply 177 * (128/255) = 89
• B_bad = ((196/255)^2.2 * (128/255)^2.2)^(1/2.2) * 255 = 98 <-- id, this is 196 * (128/255) = 98

I see two or three potential bugs, or I may not have understood what happened:

1. converting via EXR should give the same result as to PNG
2. PNG result is too dark (darker than the EXR version)
3. Even the EXR version seems wrong with respect to what I compute

This may also be the root cause of bug https://github.com/NatronGitHub/Natron/issues/582

See also:

• What every coder should know about gamma which discusses the Photoshop option “Blend RGB colors using gamma 1” (which should be set by default, but is not because of 8-bit legacy)
• [css3-images] We should use linear RGB color space when doing gradients (and transitions, and …)
• Gamma error in picture scaling
• https://discuss.pixls.us/t/ocio-srgb-alpha-conversion-problem/21040

Platform information:

• OIIO branch/version: 2.2.7 (from Mac homebrew)
• OS: macOS 11.2
• C++ compiler: from Xcode
• Any non-default build flags when you build OIIO: homebrew build

ramp.png:

solidblack.png:

rampopaque.png: