test: broker-spawn integration tests flake under host load — widen/tune the waitFor budget and reap leaked processes

Update (2026-08-10): the primary failure mode originally described here turned out to be a product race, not test-infrastructure flakiness — #623's enqueue path rewrote the job file (truncate-in-place) at the exact moment the detached worker's startup read it, killing the worker silently and leaving the job queued forever; host load only modulated the collision probability. That race is fixed in #623 (writeJobFile now writes a temp file and renames it into place, with a torn-record regression test), and the suite passes 144/144 on the host that showed the failures. This issue's remaining scope is the two test-infrastructure hardenings below, which stand on their own.

Problem

The runtime tests that spawn the full background-task chain — detached worker → workspace broker → fake codex app-server → unix-socket handshake → turn/start — gate on waitFor(..., { timeoutMs: 15000 }). On a loaded host (a VM reserving most cores, an endpoint-security agent scanning every spawned process), that chain can exceed 15 s even with no functional defect, and the test times out. Observed on 2026-08-10 while validating #623: even after the product race above was fixed, one run had a test miss the budget by ~1.2 s purely from suite parallelism plus per-spawn scanning overhead.

A compounding factor: a test run that fails or is interrupted leaks its temp-workspace broker and app-server processes (codex-plugin-test-* cwd). These accumulated to ~150 idle node processes across a few failed runs, degrading subsequent runs until manually swept with pkill -f codex-plugin-test-. A near-green run (142/143) still left 66 such processes behind, so teardown is best-effort even on the passing path.

Proposal

• Make the wait budget scalable: apply an env multiplier (e.g. CODEX_COMPANION_TEST_TIMEOUT_SCALE) inside the waitFor helper (tests/runtime.test.mjs:20) so loaded hosts and slow CI runners can widen every budget in one place, keeping the defaults for fast machines.
• Reap spawned processes on test teardown unconditionally (including timeout and passing paths) so runs cannot leak workspace processes and poison later runs.

Out of scope

Product-code changes. The torn-read race originally blamed on load is fixed in #623; the spawn chain's inherent cost is by design for these integration-level tests.

Start in tests/runtime.test.mjs at the waitFor helper around line 20, then trace how the detached worker, broker, and app-server processes are spawned and torn down. Add the configurable timeout scaling and ensure teardown runs on failures, interruption, and passing paths; done means the runtime suite tolerates slower hosts without leaked codex-plugin-test-* processes.