openscad_gui.cc's GUI-test dispatch loop iterates app.windowManager.getWindows() directly:

for (auto w : app.windowManager.getWindows()) {
  failureCount += runAllTest(w);
}

WindowManager::getWindows() returns a const QSet<MainWindow *>& — a reference to the live set, not a copy — so if any test opens a new window while this loop is running (as a real "File > New Window" test would, and as the new regression test below does), the QSet can mutate mid-iteration. That's undefined behavior for a range-for over a live container: growing a QSet can trigger a rehash that invalidates every existing iterator, including the one this loop is using.

Fix: take an explicit snapshot before iterating.

const QSet<MainWindow *> windows = app.windowManager.getWindows();
for (auto w : windows) {
  failureCount += runAllTest(w);
}

Added TestMainWindow::checkOpeningWindowDuringTestRunDoesNotCrash(), which opens a second window mid-run (new MainWindow(QStringList()) — on_fileActionNewWindow_triggered() is a private slot, not reachable from the test) and asserts the window count grew by exactly one, then closes the extra window so later tests see the expected single-window state.

This is part of the preparatory refactoring series for re-implementing #6977 (declined on size) as a reviewable stack of smaller PRs.

Testing

• Test-first attempted: with the loop unfixed, running --run-all-gui-tests repeatedly (3 runs) did not crash — with the window count only going from 1 to 2, this build's QSet bucket table doesn't happen to rehash, so there's no deterministic crash to assert on at this scale. The bug is genuine regardless (iterating a container while mutating it is UB per the standard, independent of whether a particular run's hash-table size happens not to trigger a rehash) — the new test guards the actual invariant (opening a window mid-run does not change the window count in a surprising way) and the snapshot fix removes the UB outright.
• With the fix applied: ./OpenSCAD --run-all-gui-tests — all suites pass including the new test, 0 failures.
• ctest -j8 on macOS: 3 known SVG arc tessellation failures only (same as master).
• ./scripts/beautify.sh --check clean with clang-format 18.1.8.

User documentation (for the openscad.org wiki)

No user-visible change — internal test-infrastructure fix only.

🤖 Generated with Claude Code