The set-display script needs to be sourced, not run

This removes the executable bit and shebang, names it `.sh` as is
traditional for a file being sourced, and sources it.

The reason it needs to be sourced is that it is using a symlink
that goes through an entry specific to the current proces under
`/proc`. This is done because the symlink is one of the symlinks in
the process filesystem that has special semantics: it refers to
stdout for the process, even when stdout is not a file that could
otherwise be accessed with a path on disk. (We need that because
the stdout and stderr streams in a GitHub Actions step go to an
unnamed pipe object.) But if it is run as a script then the new
shell instance that runs the script is the current process, which
then goes away, breaking the symlink (or causing it go to the wrong
place if another process is created that reuses the old PID).

Author: EliahKagan

.github/workflows/ci.yml

@@ -100,26 +100,25 @@ jobs:
           ln -s -- /usr/local/bin/wrapper1 "$base"
           EOF
 
-          # Define a script that wires up the `~/display` symlink that `wrapper1` uses to report
-          # calls as GitHub Actions step output (in addition to writing them to a log file). This
-          # is needed because stdout and stderr are both redirected elsewhere when the wrapper
-          # actually runs, and `/dev/tty` is not usable. This script must be run in the same step
-          # as the `cargo` command that causes wrapped executables to be run, because different
-          # steps write to different pipe objects. (As implemented, this also needs the calling
-          # shell to remain running, but that is not the cause of the underlying limitation.)
-          cat >/usr/local/bin/set-display <<'EOF'
-          #!/bin/sh
+          # Define a helper file that, when sourced, wires up the `~/display` symlink `wrapper1`
+          # uses to report calls as GitHub Actions step output (in addition to writing them to a
+          # log file). This is needed because stdout and stderr are both redirected elsewhere when
+          # the wrapper actually runs, and `/dev/tty` is not usable. This must be sourced in the
+          # same step as the `cargo` command that causes wrapped executables to be run, because
+          # different steps write to different pipe objects. (This also needs the shell that
+          # sourced it to remain running. But that is not the cause of the underlying limitation.)
+          cat >/usr/local/bin/set-display.sh <<'EOF'
           ln -s -- "/proc/$$/fd/1" ~/display
           EOF
 
-          chmod +x /usr/local/bin/wrapper1 /usr/local/bin/wrap1 /usr/local/bin/set-display
+          chmod +x /usr/local/bin/wrapper1 /usr/local/bin/wrap1
           mkdir /run/lock/wrapper1.fbd136bd-9b1b-448d-84a9-e18be53ae63c.lock
 
           find /usr/lib/gcc \( -name cc1 -o -name cc1plus \) \
             -print -exec /usr/local/bin/wrap1 {} \;
       - name: Build max-pure with limited dev tools and log cc1
         run: |
-          /usr/local/bin/set-display
+          . /usr/local/bin/set-display.sh
           cargo install --debug --locked --no-default-features --features max-pure --path .
       - name: Show logged C and C++ compilations (should be none)
         run: |