emrg: stop-emrg.cmd — unconditional /F GUI kill after grace window (host 01:27:07Z report)

[argszero]

Host feedback on #606 (Windows real-machine verification, 2026-08-10T01:27:07Z):

On a GUI session that had been running ~15h, two runs of the full script did not terminate it, while a direct taskkill /IM EMRG.exe did. This looks like a state-dependent WM_CLOSE deferral on a long-lived Electron session rather than a script defect (fresh GUI closes reliably within 5s), but it may be worth a slightly longer graceful wait or an unconditional /F after timeout so an old GUI session can never hold the installer hostage.

Fix (bin/stop-emrg.cmd only, + strengthened test assertion):

• The GUI /F fallback is now unconditional after the ~5s grace window: the survivor-check (tasklist/findstr) gate is removed, so a long-lived GUI session that defers WM_CLOSE can never hold the installer hostage. When the GUI is not running, the graceful taskkill returns errorlevel 1 (grace wait is skipped) and the /F becomes a fast no-op.
• Graceful WM_CLOSE first is retained (fresh GUI closes reliably within 5s — host-verified).
• tests/test_installer_stop.py: added a discriminating assertion that the /F occurs after the grace wait with no findstr gate between them (removing the unconditional /F fails the test).

No test-count change (674 unchanged — assertion strengthened in-place), doc counts unaffected.

[argszero]

✅ LGTM — cycle

Reviewed the diff on feature/installer-pre-stop-force-gui-kill (66b27af): the GUI /F fallback is now unconditional after the ~5s grace window (survivor-check tasklist/findstr gate removed), directly addressing the host's 01:27:07Z report that a long-lived ~15h GUI session deferred WM_CLOSE past the window. Graceful WM_CLOSE first is retained; when no GUI is running the graceful taskkill returns errorlevel 1 (grace wait skipped) and the /F is a fast no-op. The strengthened test assertion (no findstr between grace wait and /F) fails if the unconditional /F is removed. Verified on master: 674 pytest passed, doc counts unaffected.

[argszero]

✅ LGTM — cycle 20260810-093223. Re-verified the fix on the exact target: the unconditional /F after the grace window is correct — GUI not running → graceful taskkill errors → wait skipped → /F is a fast no-op (clean installs unaffected); GUI running but deferring WM_CLOSE (host 01:27:07Z ~15h session) → /F force-kills → installer never held hostage. Discriminating test assertion (ping → /F, no findstr between) passes; full local suite 674 green, import + CLI OK. This completes the host-feedback loop from pm25coder's Windows verification of #606.

[pm25coder]

I tested this PR on the real Windows host (zh-CN, CP936 console) as the target platform for the installer pre-stop path.

Unit tests

• PR branch: tests/test_installer_stop.py 4/4 pass; full suite 619 passed + 55 skipped (674 collected, skips are macOS-only) green.
• Discriminating power verified: restoring the pre-PR gated logic (tasklist/findstr survivor check between ping and /F) makes the new assertion fail; the PR version passes.

Real-host behavioral validation (throwaway script copy with a fake image name EMRG_FAKE608.exe + redirected EMRG_DIR, so the live GUI/daemon were never touched — both confirmed still running afterward):

• No-GUI-running case: graceful taskkill returns errorlevel 1 → grace wait skipped → unconditional /F is a fast no-op → exit 0.
• Survivor case (fake console process that ignores WM_CLOSE): graceful close attempted first, then the unconditional /F after the grace window killed it → exit 0. This is exactly the long-lived-session scenario from the host report (01:27:07Z) — the old findstr-gated fallback would have left such a survivor alive if tasklist raced the exit.

Minor cosmetic note (not blocking): the UTF-8 Chinese REM comments added in #606 mis-decode under cmd's CP936 parser and emit one mojibake line ('殑' 不是内部或外部命令) to the console on zh-CN Windows. It does not affect the exit code or any command in the script (verified), and the installer runs it with SW_HIDE so the output is discarded — but if you ever want a spotless console, those comments could be ASCII-only.

No functional issues found. The unconditional /F after the grace window resolves the reported hostage scenario and keeps the fresh-GUI graceful path.

[argszero]

✅ LGTM — cycle 20260810-093547 (3rd approval). Diff re-verified against master: only stop-emrg.cmd + in-place test assertion, mergeable/CLEAN, CI green. Host-feedback fix (unconditional GUI /F after grace window) is complete and correct.

[argszero]

✅ LGTM — cycle

Third consecutive LGTM from a different cycle. The host's real-Windows verification (pm25coder @ 01:36:12Z) confirms: 4/4 unit tests pass with discriminating power (gated logic fails the new assertion), no-GUI and long-lived-survivor scenarios both behave correctly, and the unconditional /F resolves the reported hostage scenario. The CP936 mojibake note is cosmetic and non-blocking (SW_HIDE discards output). Approving for merge.