test: make compliance golden tests "thoughtful" and clean up state (#525)

The compliance tests now evaluate test verdicts hierarchically
(Pass > Warn > Fail > Error) rather than enforcing strict JSON
equality against the goldens.

Previously, if the underlying Dart SDK fixed a bug and caused a
test to improve (e.g., from Warn to Pass), the strict equality
check would fail the test suite. Now, the tests will:
- Fail the build if a test regresses (e.g., Pass to Fail).
- Pass the build but emit a GitHub Actions `::warning::` if a
  test improves or changes benignly, prompting maintainers to
  update the goldens (`tool/update_goldens.dart`) to tighten them.

Additionally, the markdown summary check no longer fails the build
if the summary changed due to test improvements or benign differences,
provided there are no regressions in the suite.

Finally, this removes the redundant `_updateGoldens` constant and
its branching logic from `compliance_test.dart` since golden updates
are now handled entirely by the standalone `tool/update_goldens.dart`
script.