fix(tests): keep the suites inside their own temp dirs, and stop them leaking

[SUaDtL]

Two defects, one class: the test suites reach outside their sandbox.

#442 — the suite mutated the developer's real home

Running the sanctioned hook suite rewrote ~/.claude/settings.json (both statusLine.command and _codearbiterStatuslineOwner repointed at whatever plugin root the test process resolved) and littered ~/.codearbiter/ with a ledger, its lock, five session shards, and an update cache.

Four modules did it — test_session_start, test_standup, test_colorlib, test_statusline — and none used redirect_home, which already sat beside them in _helpers.py.

It broke the maintainer's statusline three times in one day. It makes the suite non-hermetic, so two developers running it end up with different global config. It is why a subagent tripped an "irreversible local destruction" check trying to clean up the mess. And CI never noticed: a fresh runner has no pre-existing settings to clobber.

The fixture is module-level, not per-class, on purpose. The leak is module-wide, those two modules carry 16 and 23 test classes between them, and a per-class setUp is one forgotten override away from regressing — while setUpModule covers every class added later for free. It pins ~ and the two user-global state seams (CODEARBITER_LEDGER, CODEARBITER_UPDATE_STATE) at one temp dir.

#462 — the suite leaked handles and processes

The same unchanged tree produced FAILED (errors=2) on one run and OK on the next three. On Windows an open handle blocks TemporaryDirectory cleanup and a live child holds a temp path, so teardown raises instead of the assertion failing — intermittent errors, not failures, with no code change between runs.

leak	fix
unclosed reads at test_wire_statusline.py:357/:604, test_init_codearbiter.py	routed through a with block
5 never-reaped subprocesses in test_taskwrite	reaped via addCleanup, so an assertion firing before _finish() can't leave a live child holding the fixture dir
test_session_start spawn handle	communicate() rather than wait() — drains and closes the pipes
implicitly reclaimed HTTPError	closed; constructed with fp=None it still allocates a spooled temp file
the two detached spawns main() fires	stubbed in _MainHarness

That last one is the interesting one. session-start.py main() ends by firing a real git fetch and a real update-refresh.py, fully detached and never awaited — correct in production, wrong in a test. Those were the five leaked pids. Both seams exist for injection; the harness now records instead of launching, which also drops test_standup from ~90s to 0.4s.

The guard

.github/scripts/test_suite_hermeticity.py, wired into the hooks job.

The obvious guard — hash the developer's real ~ before and after — is destructive by observation: it can only detect damage already done, on the machine you care about. Instead each suite runs in a subprocess whose home is a pristine temp directory seeded with a plausible stale-but-real settings.json, and that directory must come back byte-identical: nothing created, modified, or removed. Strictly stronger, and safe to run on the machine that was being damaged.

The seeded settings.json is load-bearing: heal_statusline_wiring returns early when no settings file exists, so an empty fake home would hide exactly the write this issue is about.

It deliberately does not pin the ledger/update-state env seams. Those are what a well-behaved test uses to point its own state at its own temp dir; pinning them would force correct code to look incorrect and — worse — would let an offender pass by writing to a path the guard itself supplied.

AC	covered by
AC-1 / AC-2	the seeded home is byte-identical after the run: no settings rewrite, no ~/.codearbiter/ entries
AC-3	the guard fails if any suite writes outside its temp dir — proven by two deliberate offenders, because a hermeticity guard that cannot fail is decoration
AC-4	the same audit applied to the .github/scripts suites (hook-guard matrix, codex adapter, dual-host store)

Plus a class asserting the suite emits no ResourceWarning, which is what keeps #462 closed rather than merely fixed.

Verification

Red first: the guard reported all 10 leaked paths against the unmodified suite, and 9 ResourceWarnings.

Green after: 5/5 guard tests pass, hook suite 1153 tests OK, test_ci_impact 40 OK, check_docs_contract OK, sync-core --check OK.

Closes #442. Closes #462.