[raiseCatError]

Stacked on #421. Fixes the three CI flakes seen during the v0.21 stack (plus one seen in a local release run) at their cause: no longer sleeps, reruns or looser assertions. Two of them were product defects.

Flake	Root cause	Fix	Evidence
backgroundSession.test.ts:92 (macOS, #417)	Product. When a task's leader exited, the supervisor set leftovers = group.alive, but that is the guard's own liveness. The guard is a group member and confirms the group is empty with ps before it exits. Every cleanly exiting task showed "exited, processes left" (red) for one probe, and Delete was refused.	The guard marks left when members remain. A cleanly exited leader finishes on the guard's verdict: guard exit means nothing is left, the marker means something is. A 3 s bound applies if the guard can't tell.	New supervisor test fails on the old code every run. The original test failed 3/60 under 10× parallel load, 0/60 with the fix.
configuredCompletion.test.ts:89 (Ubuntu, #417)	Product. At a startup timeout, NMSh's reset and the helper's own rm -rf remove the private root while a helper shell may still write its pid file. rmSync threw ENOTEMPTY (macOS, 39/300) or, on Linux/Node 26, returned with the directory still present (24/300). The throw also escaped a timer.	removeDirectoryNow: rename into a fresh sibling first (atomic), so no path is left to write into, then remove the copy in bounded passes. It never throws.	New race test (120 rounds of writer + remover) fails with plain rmSync on macOS and Linux and passes with the fix on both.
Ubuntu shard hung 6 min until the job limit (#415)	Test infrastructure. The reporter flushes a file when it completes. Output stopped after shellAdapters, so shellContextEnvironment never completed, and a fish was still alive at cancellation. A test that passes its timeout is cancelled, but nothing ends what it started: there, start() and its prompt waits sit outside the try/finally that kills the shell. The live PTY keeps the file's process alive and its failure is never printed. 121 files start live sessions this way.	The runner preloads scripts/test-file-teardown.mjs into each test file. Once the file's tests finish, an unref'd timer gives the file's own cleanup 15 s; if the process is still alive, every process it left running is ended (its own groups, never the runner's) and named on stderr.	Runner test: a stalled file with a piped child now fails in ~1 s with nothing left; without the preload it waits for the child.
nativeCaptureLifecycle "reaps a PTY before …" (local verify:release)	Test. Its 3 s bound also covered the owner's cold node --import=tsx compile under full-suite load.	The owner marks when it has loaded; the 3 s bound starts there.	Not reproduced in isolation (40/40 before and after); the full release run is the check.

Linux evidence came from a Debian container (Node 26.12, zsh 5.9, Fish 3.6) whose disk is on the T7. The underlying Fish stall behind the #415 hang (a prompt not arriving in time once on CI) was not reproduced: 20 Linux runs of shellAdapters and the full context-environment file passed. If it recurs, it now shows as a named test failure with the process listed.

Refs #410

[raiseCatError]

Added c350057: a fourth flake, seen once on dev itself in a dispatched CI run. codexAdapter.test.ts:84 read the fake Codex server's log the moment the turn settled. When request 101 reached the client after the person's answer to 100, the refusal of 101 arrived at the server after the turn completed, so the log was one message short. This is a test race, not a product defect. The fixture now forces that order every run (the old test then fails 3/3), and the helper waits, bounded, for the messages the test asserts (5/5 pass). Also 8f57e6a: the renamed copy in removeDirectoryNow is now removed entry by entry, because GNU rm -rf follows a renamed directory through its descriptors (Ubuntu CI on this PR). Merged forward into #423.