Stop reporting a failed comparison as a clean result

[reuvenharrison]

When oasdiff cannot load or parse a spec (exit code >= 2), the breaking and diff actions printed the error annotation and then carried straight on to write No breaking changes / No changes.

The step still failed, so CI caught it, but everything a person actually reads described the run as clean: the step output, the job summary, and the PR comment. It is also wrong for anyone using continue-on-error.

The fix

Exit as soon as the comparison fails. This is what the changelog action already does:

if [ "$exit_code" -ge 2 ]; then
    [ -s "$_err" ] && echo "::error::$(tr '\n' ' ' < "$_err")"
    if [ "$exit_code" -eq 123 ]; then
        echo "::error::…allow-external-refs remedy…"
    fi
    rm -f "$_err"
    exit "$exit_code"
fi

breaking and diff were simply missing the early exit. Nothing downstream has anything to report about a comparison that never ran, and stopping there keeps the existing 123 remedy message as the most specific thing the user sees rather than burying it under a generic failure notice.

validate had it too

Reviewing the other two actions turned up the same defect in validate, so it is fixed here rather than left for a follow-up. It shows up in the outputs instead of the text, which is worse: after the annotation it continued to a second oasdiff validate run that fails identically and prints nothing, and the count reading that output cannot distinguish it from a valid spec.

# A valid spec prints nothing, so the count stays 0.

An unloadable spec therefore published findings=0, error_count=0, warning_count=0, info_count=0. The step failed, but anything gating on those outputs reads a broken spec as a clean one, and unlike the text case that is machine-readable.

changelog is the only one of the four that was already correct.

Tests

One regression job per action, using a spec whose $ref cannot be resolved (exit 102). Each asserts the step fails and that the output never contains the clean-result text, so a future refactor cannot reintroduce the fallthrough. The validate job asserts the finding counts are absent rather than zero.

Worth noting the existing test-error-annotation.yaml already loops over all five actions at exit 102, validate included. It passed throughout, because it only checks that an annotation is emitted, which validate always did before carrying on. That is why the defect survived.

Credit

Found and reported by @lucas-monteiro-g in #198, along with the observation that the failure has to stay observable in a fixture. That PR also added comparison-status and comparison-error outputs; those are left out here since the step's own failure already signals the condition and new outputs are permanent API.