row-label-wrap intermittently crashes the engine (signal 4, cowdata.h:197) at teardown nf-cw7 ← Beads

Twinned from https://github.com/cfsh/ninefold/issues/796 by tools/beads/import_github.py's reconcile pass.

Unclaimed, unlabelled. Surfaced by probe_gate's post-merge sweep on c80b89e (#772), reproduced by hand immediately after.

row-label-wrap crashes the Godot engine outright on some runs — not an assertion failure:

``` ================================================================ handle_crash: Program crashed with signal 4 Engine version: Godot Engine v4.7.1.stable.mono.official … at: get (./core/templates/cowdata.h:197) ```

cowdata.h:197 is the bounds check in Godot's CowData::get — an out-of-range index into a Vector/Array.

It is intermittent, and the assertions are not what fails

Three consecutive standalone runs on the sandbox box, same tree, same command:

| run | outcome | |---|---| | 1 | crash, signal 4 at cowdata.h:197 | | 2 | all 5 expect_row_label / expect_focus assertions OK, then a wall of Leaked unsafe reference to object at csharp_script.cpp:179 finalize | | 3 | clean — [probe] done, exit 0, 5/5 OK |

So the probe's own assertions pass whenever it gets far enough to run them. The failure is at or near teardown, which is why it reads as a probe failure rather than a UI bug — and why it will keep being attributed to whichever merge happens to run the sweep.

⚠ It was NOT caused by the merge that surfaced it

probe_gate said *"this merge regressed main"* and named c80b89e. That merge's whole diff is .claude/settings.json, notes/chunks.md, tools/nflib/*, tools/pr/session_start.py and three test files — no games/, no shared/, no .cs, no .tscn, no tools/probes/. Nothing the engine loads. Recording that attribution as fact would send the next reader hunting through a tooling diff for a C++ crash.

Where I would start

• The leaked-reference block in run 2 is the same shutdown path, one severity down: mono finalize running against objects the scene tree has already dropped. A crash and a leak from the same teardown suggest one lifetime bug, surfacing differently by timing.
• Worth checking whether row-label-wrap's last steps leave a control focused that is freed during teardown — the probe ends on expect_row_label 4 'BASS' after a + new focus assertion.
• ⚠ Do not "fix" this by relaxing the probe. #572's note applies: the tempting repair for each of these is to weaken the assertion, and the assertions here are already passing — there is nothing to relax that would touch the actual defect.

Related

• #700 — butterbar-gate failing in the same sweeps. Different cause (host frame budget, #615's diagnostic), same symptom of arriving as "this merge regressed main".
• #598 — explore-walk fails in the sweep, passes standalone. Same *shape*, and if a third probe joins them the shared cause is probably sweep load rather than three separate bugs.
• ⚠ The sweep recorded a baseline of 2 failing at c80b89e, so both this and butterbar-gate are now encoded as expected-to-fail. Detail on #700.