emrg: installer pre-stop — gracefully close GUI/TUI/daemon before overwriting (rant 2026-08-10T08:50:44)

[argszero]

Windows installs of v0.2.10/v0.2.11 got stuck at "停止已有进程" (stopping existing processes) and the host had to reboot. Root cause: the Inno Setup installer overwrites %USERPROFILE%\.emrg\install (the daemon's runtime dir) without stopping running processes — the windowless pythonw.exe -m emrg.server daemon holds file locks, and Inno CloseApplications cannot see windowless processes.

Fix (pre-install graceful shutdown, order GUI → TUI → daemon):

• new bin/stop-emrg.cmd: ① GUI taskkill /IM EMRG.exe (graceful WM_CLOSE, ~5s, then /F fallback) ② TUI python.exe -m emrg via command-line filter (wmic %%-escaped LIKE, PowerShell CIM fallback for Win11 24H2+ where wmic is removed; daemon pythonw -m emrg.server excluded) ③ daemon: emrg server stop protocol shutdown via the old install's CLI (present since emrg: Phase 4 packaging 资产 + daemon ping 响应补 type 字段（rant #12 §3/§6/§10/§12/§13） #364 — version-safe), emrgd.pid poll (≤10s), then taskkill /F /PID fallback. Returns 0 when nothing survives, 1 otherwise (installer aborts with a clear message instead of hanging). Clean installs skip everything.
• bin/emrgd.cmd: new stop branch → python -m emrg server stop (reuses the tested _stop_daemon path; label flow so %errorlevel% expands after the child exits).
• packaging/make-installer.sh: .iss gains [Files] dontcopy entry + [Code] PrepareToInstall that extracts and runs stop-emrg.cmd (via {cmd}, SW_HIDE) before file overwrite; non-zero exit aborts with actionable message.
• packaging/build-runtime.sh: bundles stop-emrg.cmd into runtime bin/.
• +4 tests (tests/test_installer_stop.py, text-assertion wiring — no iscc/cmd on macOS/CI): stop-emrg.cmd step coverage + order, emrgd.cmd stop branch, .iss PrepareToInstall/dontcopy/SW_HIDE, build-runtime copy. Doc counts synced: 670 → 674 pytest (README.md / Agent.md / README.cn.md).

Verification on this host: 674 pytest green, import + CLI checks green. Windows behavior must be validated by the host on the next release (expected per the new status rules, feedback flows back via new rants).

[argszero]

✅ LGTM — cycle

Reviewed the full diff on branch feature/installer-pre-stop and verified locally:

• Full suite: 674 passed (4 new tests in tests/test_installer_stop.py green), doc counts synced (README.md / README.cn.md / Agent.md 670→674), import + CLI checks green.
• bin/stop-emrg.cmd: GUI graceful WM_CLOSE → /F fallback, TUI command-line filter (wmic %% escaped LIKE + PowerShell CIM fallback, daemon pythonw -m emrg.server excluded), daemon protocol stop via old-install CLI + emrgd.pid poll (≤10s) + taskkill /F fallback; clean-install safe (skips to exit 0).
• bin/emrgd.cmd stop branch: label flow avoids %errorlevel% expansion trap; 'emrg server stop' confirmed present in emrg/main.py (line 62/104/158 _stop_daemon).
• make-installer.sh: [Files] dontcopy + [Code] PrepareToInstall via {cmd} SW_HIDE, non-zero exit aborts with actionable message — consistent with the #592 no-console-window discipline and Inno CloseApplications blind spot for windowless pythonw.
• build-runtime.sh bundles stop-emrg.cmd into runtime bin/.

Windows real-machine validation flows back via new rants per PR #605's updated status rules.

[argszero]

✅ LGTM — cycle 20260810-091234. Re-reviewed the batch script and found + fixed a latent cmd.exe delayed-expansion bug in the :verify block (parse-time %DPID% inside a parenthesized block would make the daemon-alive check always false → false 'clean' exit). Now uses setlocal enabledelayedexpansion + !DPID!, with a discriminating regression assertion in tests/test_installer_stop.py. Local suite: 674 pytest green, import + CLI checks pass.

[argszero]

✅ LGTM — cycle 20260810-091736 (3rd approval). Final diff re-verified: delayed-expansion fix present, PrepareToInstall/dontcopy wiring intact, doc counts synced (674). CI green on updated branch.

[pm25coder]

I tested this PR on the Windows host (the exact target platform) after it was merged, and the Windows-specific logic checks out:

• tests/test_installer_stop.py: 4/4 pass on real Windows; from emrg.client.app import run_client import + emrg --help CLI checks green.
• Version-safe daemon stop claim holds: emrg server stop is confirmed present in master's emrg/__main__.py (since emrg: Phase 4 packaging 资产 + daemon ping 响应补 type 字段（rant #12 §3/§6/§10/§12/§13） #364), so calling the old install's emrg.cmd server stop is safe across versions.
• Daemon-kill path verified end-to-end on a throwaway copy (home dir redirected so no live processes were touched): the script reads emrgd.pid, force-kills the PID after the ≤10s poll, and the :verify block returns 0 when nothing survives and 1 while a process remains — the delayed-expansion fix (setlocal enabledelayedexpansion + !DPID! in the parenthesized block) works as the regression test asserts.
• GUI-kill path verified with tracing: with a freshly launched GUI (4 EMRG.exe processes), taskkill /IM EMRG.exe (graceful WM_CLOSE) closed the entire Electron tree within the 5s window — the after-check tasklist/findstr returned not-found, so the /F fallback was correctly skipped and the final process count was 0.

One observation worth carrying into the host-side validation: on a GUI session that had been running ~15h (started the previous evening), 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.

Full local verification was done on master bf2012e (the merge commit) with a clean tree.