Import #776's teardown half: decide on the root's liveness, not on taskkill's message

[Edo771977]

One new PR upstream since the last sweep. It bundles three changes; one is worth having here, and the other two are not.

347/347, full suite run twice. tsc clean.

What was wrong here

Our Windows branch trusted taskkill's exit status and, failing that, matched its message. The localization hole the PR leads with — looksLikeMissingProcessMessage only recognising English, so a French "introuvable" fell through — was already closed here: we carry the status === 128 check, which is language-independent.

The part that was broken is the one their commit message buries: any other non-zero status threw at the caller, and taskkill produces one routinely. /T walks the tree and then terminates each entry, so a short-lived git or cmd helper that exits in between makes taskkill report a failure although the root did die. A throw out of terminateProcessTree() lands in teardown paths that cannot tell that apart from a real failure.

Now:

• the root is probed before taskkill is spawned at all — a root that is provably gone returns {attempted: false, delivered: false, method: null} and costs no process
• after a non-zero status, the root's own liveness decides
• delivered accordingly describes the root, exactly like the process-group SIGTERM on other platforms. It never proved every descendant was gone, and now says so

Changes on top of the import

• their isProcessAlive() is dropped for this fork's isPidAlive(), and the new gate is isProvablyGone(): only ESRCH proves absence. Their reading (anything but ESRCH is alive) is the right one for "should I kill"; the question at this gate is the inverse — "may I skip the kill" — and skipping it on a root we merely cannot read leaks the whole tree
• two of our tests now inject killImpl. Without it they call the real process.kill(1234, 0), so their outcome depended on whether pid 1234 happened to exist on the machine running the suite. That is exactly the class of environment-dependent test this fork has spent two rounds removing

Not taken

Their change	Why not
path.posix.join in createBrokerEndpoint	the unix: branch is unreachable on Windows, so on every platform this fork runs it produces the same string
2s timeout in sendBrokerShutdown	ours is already bounded, with refused/unreachable semantics and a retry — strictly ahead of it

Their two test-portability fixes are kept, though: a .cmd shim instead of a symlink for the fake node, and a skip when Windows refuses to create the broken-symlink fixture without Developer Mode. They change no behavior and let the suite run on a real Windows host — which is the one check this fork still cannot do from here.

Sabotage matrix

sabotage	which tests fail
pre-kill probe removed	4, including "skips taskkill when the Windows process is already absent"
back to matching the message instead of liveness	2, including "uses liveness instead of localized taskkill output"
a permission error read as absence	2, including "does not treat a Windows preflight permission error as missing"

🤖 Generated with Claude Code

https://claude.ai/code/session_01UXfvnjSC72HsM6EEPVt2Tg

---

Generated by Claude Code