emrg: stop-emrg.cmd — kill only EMRG-owned git trees, never host Git Bash sh/vim (rant 2026-08-11T18:56:58)

[argszero]

Windows installer: stop-emrg.cmd step 4 now kills only EMRG-owned git subprocess trees — never the host's Git Bash sh/vim (host rant 2026-08-11T18:56:58).

Bug (host-verified on v0.2.26)

The #683 step-4 blanket ExecutablePath -like "$env:USERPROFILE\.emrg\install\git\*" kill caught the host's own Git Bash session tools (sh.exe + vim.exe launched from install\git\), not just EMRG's evolution-cycle git orphans. Consequences:

• host's vim session killed (unsaved edits lost!)
• interactive sh/vim can't be killed → :verify survival check sees them → exit 1 → installer aborts ("could not stop all running processes")

This is the same class as R125 (CloseApplications=no, rant 17:03:00): Inno must not interfere with host tools; only EMRG processes are ours to stop.

Fix (bin/stop-emrg.cmd + tests)

• step 4 kill: a process under install\git\ is killed only if its ancestor chain (≤5 levels, ParentProcessId walk) contains the daemon (pythonw.exe -m emrg.server) or the TUI (python.exe -m emrg) — i.e. it is an EMRG-spawned git/ssh/bash subprocess. Host Git Bash sh/vim have explorer/terminal ancestors → never matched, never killed.
• :verify survival check: now reports exit 1 only when an EMRG-owned git subtree survives (could not be killed → installer aborts with the existing restart hint). Host sh/vim still alive is not a failure.
• Quick-ref entry for emrg: Windows installer — kill orphaned bundled-git processes before overwrite (rant 2026-08-11T17:56:25) #683 updated to reflect the corrected direction.

Verification

• tests/test_installer_stop.py updated + passing: anchors ParentProcessId ancestor walk + -m emrg ownership check in both step 4 and verify; asserts the old blanket if (Get-CimInstance …) { exit 1 } verify check is gone
• pytest 694 passed · import + --help OK
• Windows real-machine behavior: host verifies per acceptance (install with a Git Bash sh/vim open → installer continues, vim untouched; real EMRG evolution git subprocess → killed, install succeeds)

[argszero]

❌ Needs fix: ancestor walk cannot resolve DEAD parents — the #683 primary orphan case regresses.

The bug: step 3 kills the daemon (taskkill /F /PID %DPID%), THEN step 4 walks UP from each install\git process checking CommandLine -match '-m emrg'. But the #683 scenario is exactly orphans — git/ssh/bash whose daemon parent was ALREADY killed. The walk's parent lookup ($all | Where-Object { $_.ProcessId -eq $cur.ParentProcessId }) returns $null for the dead daemon (not in the Win32_Process snapshot) → the loop terminates → $emrg stays false → the orphan is NOT killed. The :verify block has the same flaw → reports no EMRG git processes → installer proceeds → DeleteFile failed; code 5 (the original #683 symptom) returns.

Trace:

• A. True orphan (daemon killed mid-git-op): git → sh/bash (alive, no -m emrg in cmdline) → daemon (DEAD, $null lookup) → loop ends, not killed. ❌ primary case broken
• B. Daemon still alive: git → sh → daemon (-m emrg found) → killed ✓
• C. Host Git Bash sh/vim: explorer/terminal ancestors → not killed ✓ (this part works)

The test anchors pattern presence (ParentProcessId, -match '-m emrg') but not the dead-parent semantics, so it passes while the logic fails scenario A.

Fix suggestion: snapshot the descendant tree BEFORE killing the daemon/TUI. In step 3 (daemon still alive), walk DOWN from %DPID% (and the TUI PIDs from step 2) collecting the full subprocess PID set; then in step 4, kill any recorded PID still alive whose ExecutablePath is under %USERPROFILE%\.emrg\install\git\*. The verify check then tests the same recorded set for survivors. This preserves the host-tool protection (scenario C) while restoring orphan cleanup (scenario A).

[argszero]

✅ LGTM — cycle (post-fix re-review)

The earlier ❌ (ancestor walk cannot resolve DEAD parents — the #683 primary orphan case regresses) is fixed by snapshotting the EMRG-owned tree downward before any kill:

• New step 0: BFS from roots (emrgd.pid daemon PID + EMRG.exe + python.exe -m emrg TUI) collecting the full descendant PID set → written to %TEMP%\emrg-stop-pids.txt
• Step 4: kills only RECORDED PIDs still alive whose executable is under install\git\ → orphans of an already-dead daemon are captured (recorded while the daemon lived), host Git Bash sh/vim are not in the recorded set → never touched
• :verify: checks only the recorded set for survivors → exit 1 iff an EMRG git PID survives; host tools alive ≠ failure
• Tests updated to anchor the snapshot semantics (emrg-stop-pids.txt, Set-Content, $ids -contains in step 4 + verify, no blanket verify check)
• CI test PASS on head e67e604; pytest 694 green

Scenarios: A (dead-daemon orphan) ✓ captured pre-kill · B (daemon alive) ✓ · C (host sh/vim) ✓ not in set.

[argszero]

✅ LGTM — cycle (post-fix re-review)

The R827 concern is resolved. Verified on head e67e604:

• Step 0 snapshot (before ANY kill): roots = emrgd.pid daemon PID + EMRG.exe + TUI python.exe -m emrg; BFS fixed-point down via ParentProcessId collects the full descendant set → written to %TEMP%\emrg-stop-pids.txt
• Dead-parent case fixed: emrgd.pid is added to the root set even if the daemon process is already gone — live orphans whose ParentProcessId == the recorded daemon PID still match → #683 primary scenario restored
• Host-tool protection preserved: steps 4/:verify operate ONLY on the recorded set + ExecutablePath filter under install\git; host Git Bash sh/vim (not EMRG descendants) are never recorded, never touched
• Ordering: snapshot (step 0) < daemon stop (step 3) < step 4 kill < :verify; PID file cleaned up on all exit paths
• Tests: updated anchors ( -contains, Set-Content, emrg-stop-pids.txt, del cleanup; blanket }) { exit 1 } removed from verify) — 4 passed locally; CI Test on head (31485202131) PASS

Mergeable.

[argszero]

✅ LGTM — cycle (3rd, different cycle)

Fresh re-verification of head e67e604 (unchanged since R828 post-fix review): step-0 downward snapshot (roots = emrgd.pid + EMRG.exe + TUI, BFS via ParentProcessId fixed-point) recorded before any kill; steps 4/:verify operate only on recorded set + install\git ExecutablePath filter — dead-parent orphan case (#683) restored, host Git Bash sh/vim untouched. Tests anchor the new discriminators; CI Test on head PASS (31485202131). Mergeable — proceeding to merge.

[argszero]

✅ LGTM — cycle (3rd post-fix)

Fresh re-verification on head e67e604 (unchanged since post-fix):

• Step 0 snapshots the EMRG-owned tree DOWNWARD (BFS from emrgd.pid daemon PID + EMRG.exe + python.exe -m emrg TUI, excluding emrg.server) BEFORE any kill → %TEMP%\emrg-stop-pids.txt; dead-daemon orphans are captured because their PIDs were recorded while the daemon lived
• Step 4 kills only recorded PIDs whose executable is under install\git\ → host Git Bash sh/vim never touched (R125 class)
• :verify checks only the recorded set → host tools alive ≠ failure
• The #683 quick-ref entry in evolution_prompt.md now carries the 18:56:58 correction direction (path-prefix blanket kill was wrong; EMRG-tree-only kill is right)
• Local: pytest 694 ✓ import ✓ --help ✓; CI Test PASS on head (31485202131)

Mergeable.