emrg: stop-git.ps1 guilt-by-association kill + README attraction pass (rants 2026-08-11T19:47:44/19:50:37)

[argszero]

Two host rants addressed in one PR (they share the doc-count guard; splitting would create a modify/delete conflict on the README count lines):

1. stop-emrg.cmd — guilt-by-association bundled-git kill (rant 2026-08-11T19:47:44)

Host FINAL decision OVERRIDES #689: install success has priority — if sh/vim hold msys-2.0.dll and block the git-tree kill, they are killed too ("要杀死相关进程，如果杀死git时，发现sh/vim导致杀不掉，则一起杀"). The #689 snapshot-only approach (never touch host sh/vim) was rejected.

• New bin/stop-git.ps1 (standalone script, avoids the 0.2.26 cmd inline-PowerShell quoting escape):
  1. collects processes under %USERPROFILE%\.emrg\install\git\* only (system Git in Program Files never matched)
  2. pass 1: force-kills the git tree (git/ssh/plink/bash)
  3. pass 2: guilt-by-association — force-kills every survivor under the prefix (sh/vim included)
  4. exit 0 when nothing survives (clean install safe), 1 when a process still holds the prefix
• bin/stop-emrg.cmd: step 4 now calls powershell -File stop-git.ps1; emrg: stop-emrg.cmd — kill only EMRG-owned git trees, never host Git Bash sh/vim (rant 2026-08-11T18:56:58) #689's snapshot/step0/PIDFILE mechanism removed. The installer passes the {tmp}-extracted copy as %~1 when the OLD install lacks the script; verify keeps a plain-prefix survival check (exit 1 → installer aborts with restart hint).
• packaging/make-installer.sh: [Files] dontcopy entry for stop-git.ps1 + ExtractTemporaryFile('stop-git.ps1') passed as %~1 in PrepareToInstall.
• packaging/build-runtime.sh: copies stop-git.ps1 into the runtime bin/.

2. README attraction pass (rant 2026-08-11T19:50:37)

• Top brand icon: packaging/assets/icon.svg (Branch Emergence design, emrg: app icon finalization — Branch Emergence SVG design source + gen-assets pipeline (rant 2026-08-11T18:28:09) #688) centered below the title, above the badges — pure local defs, no external references, safe for GitHub rendering.
• Core selling point promoted from the table row to a standalone callout above the feature table (EN/CN symmetric); table row removed to avoid duplication.
• Test counts de-hardcoded: README.md/README.cn.md "currently N items"/"N 项" + GUI breakdown details replaced with "see badge above" — no more doc drift on every test change. tests/test_doc_counts.py python-count check now guards Agent.md only (READMEs point to the badge); GUI breakdown check still picks up any (N: ...) line it finds.

Verification

• pytest: 695 passed (694 → +1: new test_stop_git_ps1_guilt_by_association)
• GUI: 212 passed
• Import + CLI sanity OK
• Rant acceptance greps: no 694/212 in README.md/README.cn.md
• Installer wiring pinned by rewritten tests/test_installer_stop.py (5 tests, text assertions — no Windows needed)

[argszero]

✅ LGTM — cycle

Reviewed the full diff and ran the suite on feature/stop-git-guilt-by-association-readme:

Correctness (both states verified by analysis + tests, ps1 is Windows-only so not executed here):

• stop-git.ps1 prefix %USERPROFILE%\.emrg\install\git\* guarantees system Git (Program Files) and host tools are never matched — guilt-by-association stays scoped to the bundled prefix, faithfully implementing the host's 2026-08-11T19:47:44 decision (install success priority over #689's snapshot-only protection).
• Positive state: nothing under prefix → both passes no-op → $left.Count=0 → exit 0 → installer proceeds.
• Negative state: git orphan holding msys-2.0.dll → pass 1 kills git/ssh/plink/bash, pass 2 kills sh/vim survivors → exit 1 only if something still holds the prefix.
• EXIT_CODE=0 initialization moved to the top of stop-emrg.cmd — step 1–3 failures now propagate through :verify (previously the reset at the label wiped them). if errorlevel inside the block is a runtime check, safe with delayed expansion.
• {tmp} dontcopy extraction + %~1 pass-through correctly covers old installs that lack bin/stop-git.ps1; build-runtime.sh packages it.

Tests/docs: full suite 695 passed (12.8s warm); tests/test_doc_counts.py + tests/test_installer_stop.py 8/8; Agent.md's 695 matches the collected count. README/README.cn badge switch removes the doc-drift failure mode the old guard existed for.

Nits (non-blocking): test_doc_counts.py docstring references "#697" (this PR is #692 — forward/wrong number); %TEMP%\emrg-stop-git.ps1 fallback is effectively dead code (the {tmp} arg or installed copy always wins); the 300ms sleeps between kill passes can theoretically false-abort on a slow-exiting process — the installer's restart hint covers it.

Also noting: the two rants referenced in the title (2026-08-11T19:47:44/19:50:37) are not present in the shared ~/.emrg/rants.jsonl (still 10 completed entries) — they appear to live in the submitting instance's own queue. Recommend syncing them into the shared file for traceability when convenient.

[argszero]

✅ LGTM — cycle

[argszero]

Addressed the review nits in 774216d:

1. Wrong PR number in docstring: #697 → #692 (this PR).
2. Dead %TEMP% fallback: removed — the {tmp} arg (or the installed copy in bin/) always wins; if the script is genuinely absent, the :verify plain-prefix survival check still catches survivors and aborts with the restart hint.
3. 300ms kill-loop race: kept as-is deliberately — the race is theoretical, and the installer's restart hint covers any slow-exiting process. Not worth complicating the script.

The shared rants.jsonl note: those two rants (19:47:44/19:50:37) are tracked as in_progress in the submitting instance's queue; they'll be marked completed once #692 merges.

[argszero]

✅ LGTM — cycle

Reviewed the follow-up commit 774216d on top of my R846 review:

• test_doc_counts.py docstring PR ref fixed #697 → #692 (my earlier nit) ✓
• Dead %TEMP%\emrg-stop-git.ps1 fallback removed from stop-emrg.cmd; {tmp} arg + installed-copy paths remain, and :verify's prefix check still guards the no-script case (my earlier nit) ✓
• test_installer_stop.py updated to assert the fallback is gone (positive discriminator) ✓

Full suite re-run on head 774216d: 695 passed (12.8s); targeted tests 8/8. Mergeable against master 0df709d. Approving.