emrg: the GUI state test leaves three temp directories behind on every run #1916

What is wrong

emrg/gui/test/gui-state.test.js creates a temp home per calling test and never removes it, so every GUI harness run leaves three directories and ~2.7 KB in os.tmpdir() for good. mkdtemp() never reuses a name, so nothing ever reclaims them: the count grows by three per run, without bound, on every developer machine and every CI checkout that runs the harness.

Measured on this host (macOS), os.tmpdir() = /var/folders/…/T:

36 stale `emrg-gui-state-*` directories / 96 KB      (before this cycle)
+3 after one `cd emrg/gui && npm test`               → 39
+3 after the next run                                → 42

Scope — it is this one file, and the measurement says so

Four files under emrg/gui/test/ call fs.mkdtempSync:

daemon_client.test.js:86,198,209    integration.test.js:63    conn-manager.test.js:54    gui-state.test.js:11

Per-prefix counts taken immediately before and after one full harness run — only one prefix moves:

prefix              before   after
emrg-gui-test-           0       0
emrg-gui-state-         39      42     ← the only growth
emrg-gui-integ-          0       0
emrg-conn-test-          0       0
emrg-venv-               0       0

The other three already hand their directories back — daemon_client.test.js:103 and conn-manager.test.js:72 remove theirs in a teardown, integration.test.js:167 has an after() hook. So this is not a new convention being proposed: it is one file out of step with the three beside it in its own directory.

Why it matters beyond tidiness

A test that creates host state owes the removal of it, exactly as it owes its assertions. The failure mode is quiet by construction — a green run and a lengthening ls of os.tmpdir() look identical — so nothing in CI reports it, and the cost is paid on machines that run the suite most.

Acceptance

• After cd emrg/gui && npm test, the count of emrg-gui-state-* directories under os.tmpdir() is unchanged by the run.
• The cleanup is asserted rather than assumed: an arm that removes the hook must go red (a silently failing removal would otherwise be the same litter with a tidier story).
• No other test file's temp handling changes — the three above are already correct and are left alone.

Named limit

Only the emrg-gui-state-* prefix is asserted. The Python suite's temp handling and the landing-tree gates' worktrees are a different surface and are not claimed here.