Could you post an example of the resulting wrapper before and after this change?

Before:

export NIX_CFLAGS_COMPILE=-isystem /nix/store/3cv0q9g9zyg5alzzz6072rhdmvzhk5yy-cuda-floatn.h/include${NIX_CFLAGS_COMPILE:+ }$NIX_CFLAGS_COMPILE

After:

export NIX_CFLAGS_COMPILE="-isystem /nix/store/3cv0q9g9zyg5alzzz6072rhdmvzhk5yy-cuda-floatn.h/include${NIX_CFLAGS_COMPILE:+ }$NIX_CFLAGS_COMPILE"

This swaps no quoting for approximately correct quoting. Please do not add double quotes, use "${value@Q}" and "${separator@Q}" instead of $value and $separator. The result will look like this:

export NIX_CFLAGS_COMPILE='-isystem /nix/store/3cv0q9g9zyg5alzzz6072rhdmvzhk5yy-cuda-floatn.h/include'${NIX_CFLAGS_COMPILE:+' '}$NIX_CFLAGS_COMPILE

Excellent, thank you for the catch! To be honest I didn't know about @Q Bash feature before, it makes many things which I have in mind simpler.

Can you review this version? It passes my new test for me with the following wrapper:

export FOO='ahaha $ \'
export BAR='ahaha $ \'
export BAZ='ahaha $ \'
export QUZ='ahaha $ \'
export FOO='bar '\''$baz'${FOO:+' '}$FOO
export BAR=$BAR${BAR:+'' ''}'bar '\''$baz'
export BAZ=$'test1\n test3\ntest4'${BAZ:+$}$BAZ
export QUZ=$QUZ${QUZ:+$}$'test1\n test3\ntest4'

@grahamc Bot forgot to remove old labels after a force-push.

How can I quote @Q in a commit message so that to not ping @Q the member with it? @q, sorry for the noise!

EDIT: fixed it with a rewording for now.

To be honest I didn't know about @Q Bash feature before, it makes many things which I have in mind simpler.

I've learnt about it not long ago; before that I'd have suggested to use $(printf %q "$var") which achieves the same a bit more verbosely.

I have set up a Hydra cluster and tested this. Thank you!

Somehow I missed some failures… This is causing a bit of trouble in staging in cases such as --set CC "$""{CC:-$CC""}" and --set RETROFE_PATH "\${RETROFE_PATH:-\$PWD}". I'm going to introduce --set-default for the first kind of usage (set to a constant unless already set) and --set-eval for the latter (set to an arbitrary expression).
/cc @vcunat

The docs refer to the source files for such details, so I see no direct need to update the manual. It seems too low-level detail for release notes, too, not affecting any "normal" uses of the wrappers.

I've reviewed all usages of makeWrapper / wrapProgram in Nixpkgs and found 6 additional expressions that had to be updated in fb703ad (and eb3d207). It is not unlikely that when this reaches master there will be a couple runtime breakages invisible on Hydra and missed during the manual review.

Well that explains why a few of my wrappers suddenly stopped working..