There must be a separate paragraph about "\s0" and mentioning of 0
for \s+-N dropped.

  Adding text to document the use of "\s0":

\s0        Restores the previous type size.

        Does not restore the intended type size if there is a type size
        shift before it, like "\s<i>...\s<n>...\s0...\s0" (restores
        \s<i>+<n>, not that before \s<i>).

        The best practice is not to use it.

        The second best practice is not to use it after a variable,
        as it might contain "\s0".

There's nothing special about \s in this regard.  Many requests and escapes take a special argument (or use lack of one) to indicate "go back to the previous value."  In isolated cases (one, and maybe the only, example being the .ev request), roff maintains an internal stack, so the "go back" syntax actually can be nested.  But this is the exception rather than the rule.  Typically "go back to the previous value" does exactly that, and while there are plenty of aspects of roff syntax lying in wait to trip up the unwary, I think the documentation makes this behavior pretty clear; a rudimentary understanding of "previous value" will clue in the reader that what coders think of as nesting won't work.

Certainly there is more risk of getting tripped up by this when a macro calls another macro, or a string another string, without understanding all state changes that macro or string might make, and this could be a point to make (if it's not there already) when discussing a high-level view of the language.  Do you have a proposed wording for this?

Subject: [doc]: Add a remark about a pitfall using \s0

  Using two type size changes around a string can cause malformed
changes if the string (with the size changes) is surrounded by a size
change.

  Example from gcc(1), which uses (pod2man):

.ds C+ C\v'-.1v'\h'-1p'\s-2+\h'-1p'+\s0\v'.1v'\h'-1p'

  or simplified

.ds C+ C\s-2+\s0

  and uses

\s-1ISO \*(C+\s0

  So it is actually:

\s-1...\s-2...\s0...\s0

  which results in a sum of \s-1.

  So, when defining a string that contains local type size changes
(\s), absolute size changes should be used (\s+n...\s-n; \s-n...\s+n),
not \s0.

  See also bug #40793.