Windows: broker child processes leak and stale broker.json poisons setup tests on persistent machines

Two related Windows process-hygiene gaps found while running the suite repeatedly on a persistent (non-CI) Windows 11 box. Both are pre-existing (independent of any local change) and reproducible.

1. Stale broker.json poisons setup-family tests

%TEMP%\codex-companion\<repo-hash>\broker.json persists across runs and can point at a dead broker PID / stale named pipe. The setup-family tests key shared state on the real repo path (the test fixture's buildEnv() inherits process.env unmodified rather than using an ephemeral fixture dir), so a stale pointer makes every subsequent run report false "not ready" results. Symptom: ~6 spurious failures that vanish permanently after deleting that one directory. On ephemeral CI this is masked because the temp dir is always fresh.

Suggested fix: have the setup tests use an isolated $TEMP/HOME per run, or validate/ignore a broker.json whose PID is dead before treating it as authoritative.

2. Broker child processes leak on Windows

Repeated test runs accumulate orphaned app-server-broker.mjs node children (observed ~200 after a day of runs), each holding its temp workspace open. Same family as the known taskkill /T /F flake ("operation attempted is not supported" on a nested child). Correctness impact is low (each is isolated to a random temp workspace), but it blocks git worktree remove on any worktree a broker still holds open, and exhausts the process table over time.

Environment: Windows 11, codex-cli 0.144.1, plugin at current main. Happy to provide fuller repro details.

Start with the setup-family tests and their buildEnv() path, then inspect how they consume %TEMP%\codex-companion<repo-hash>\broker.json. Trace cleanup for app-server-broker.mjs, including the existing taskkill /T /F handling. Done means repeated Windows runs neither trust dead broker state nor leave broker child processes holding workspaces open.