> The origin of the faces without a foreground setting is commit
> 382ceb2cdbedc06b06d9fb424d83f531339a3311 by Juri Linkov <juri@jurta.org>
> while working on https://debbugs.gnu.org/cgi/bugreport.cgi?bug=10181.
>
> I've cc'd him here. Juri, can you comment?

Thanks for pointing me at this issue since I had not realized
this is related to something affected by my previous changes.

> I'm especially interested to hear your thoughts on why the change made
> things nicer (what you said in
> https://debbugs.gnu.org/cgi/bugreport.cgi?bug=10181#84).

The core point of my 2014-year message was that with the specified foreground,
Ediff removes font-lock highlighting on non-current differences.
The need for keeping the font-lock foreground colors is evident
when looking at these font-lock colors in buffers where the effect
of font-lock is noticeable, such as changes in the source code files.

But I completely agree that the current contrast is too low on a dark theme.

Fortunately, now we have a feature that will help to fix this problem:
the face attribute :distant-foreground specifies the alternative
foreground color that is used only when the background color is too
close to the foreground color, thus making the text readable in this case.

> Or, perhaps the ediff faces should vary the background based on
> (background light) and (background dark)?

The background colors could be adjusted as well for dark backgrounds.

After changing background colors to darker colors and adding
:distant-foreground I tried different default foreground colors with

  M-x set-foreground-color RET dark grey RET

and it seems everything looks readable.

Do you see more problems with the following patch?