install-whitaker's test harness runs fragments with bash -c, which is not how a runner runs them #449

Problem

install-whitaker's test harness does not invoke a composite fragment the way a
runner does, and the difference is observable on the macOS runner image.

A runner writes a composite run: fragment to a file and executes that file.
.github/actions/install-whitaker/tests/_fragment_runner.py:210 passes the same
text to bash -c instead:

    process = subprocess.run(  # noqa: S603,TID251 - exercise the Bash fragment.
        [bash_executable(), "-c", script],

Those are not equivalent. Bash 3.2, which macOS runners ship, replaces itself
with the last command of a bash -c string when that command is external,
discarding any ERR trap along with the shell. Bash 5 suppresses that
optimization when a trap is set.

Evidence

Six probes on macos-15 and ubuntu-latest, differing only in what fails,
where, and how the script is invoked. The macOS shell is
GNU bash, version 3.2.57(1)-release (arm64-apple-darwin24) at /bin/bash;
the Linux shell is 5.2.21(1)-release.

So the only affected shape is a bash -c string whose last command is
external, on Bash 3.2. A file is never affected, which is why a runner is never
affected.

Impact today: latent, not live

install-whitaker reports seven failure metrics from ERR traps, and its tests
assert several of them (test_install_whitaker_install.py,
test_install_whitaker_windows_paths.py). Those assertions are sound at
present: no install-whitaker fragment ends in a bare external command. Their
last commands are builtins, function calls, or redirections, none of which the
optimization applies to.

The gap is that nothing keeps it that way. A fragment that grows an external
command at its end would, on the macOS leg of python-tests, report a failure
with no ::error annotation and no metric, and the test asserting that metric
would fail for a reason that has nothing to do with the action. That is not
hypothetical: it is exactly what happened while #443 was being written, and it
cost a wrong diagnosis, a manifest test that banned ERR traps on false
grounds, and a retracted claim that Bash 3.2 does not run traps at all.

Proposal

Fold install-whitaker's private harness onto the shared
composite_fragments.py added in #443, which writes each fragment to a file and
runs the file, and which carries two regression tests for exactly this:

• one asserting a fragment's $0 names a file rather than bash;
• one asserting a fragment's ERR trap fires when its last command fails.

install-whitaker's copy predates the shared module and is otherwise the same
code, so this is a deletion and an import, plus whatever small differences the
two have drifted into. Doing it also removes a 240-line duplicate.

No change to install-whitaker's ERR traps is needed. They work on a runner.

Not affected

install-nixie uses no ERR traps.

Notes for whoever takes this

It touches .github/actions/install-whitaker/tests/, which had in-flight work
in #448. Rebase onto that first.