[gregoryfoster]

Closes #40. Test-only: no deploy (tests/ isn't a runtime path).

The strand test: no wall clock

The hand-off's trace confirmed. With a 0.6 s in-process extraction, the old test fails with ['cmd-1', 'cmd-1', 'cmd-2']: step 1 took longer than 500 ms, so the second walk reclaimed cmd-1 again. It never failed in 30 runs under CPU load, so the flake is rare.

Now XCLAIM … IDLE 60000 JUSTID, run as admin, gives both entries 60 s of idle time. They stay with the dead consumer, and JUSTID means no delivery is counted. The threshold is reclaim_min_idle_ms=30_000, and the sleep(0.6) is gone. Scratch Redis and CI are both 7.0.15, which accept IDLE.

• It passes with a 2 s slow step 1, where the old test broke at 0.6 s.
• It still proves its point. With the mutation count=1 → count=2 in Consumer.reclaim (claim both entries at once), it fails ['cmd-1'] == ['cmd-1', 'cmd-2'].

Sweep of tests/

Fixed:

Test	Margin	Fix
test_drift.py::test_every_call_shares_one_deadline	1.5×: a 0.2 s delay against a real 0.3 s deadline. Overhead over 0.1 s fails the first call	A fake monotonic for the deadline. The second call gets 0.1 s against a 0.5 s stall (5×), and the first call has no timing in it. Mutation timeout=REQUEST_TIMEOUT_SECONDS (no shared deadline) → fails
test_deploy.py::test_refuses_while_another_deploy_holds_the_lock	A race, not a margin: it polled for the lock file, but flock(1) creates the file before it holds the lock	The test holds the lock itself (fcntl.flock). Mutation: deploy.sh skips flock -n → fails

Listed, safe (≥ 2× or one-sided):

Site	Why safe
test_consumer.py test_run_stops_promptly: sleep(0.5)	Only lets the loop run. Any length passes. The stop bound is 5 s against a 100 ms block (50×)
test_consumer.py broker-name retry: poll 10 ms, asyncio.timeout(10)	A bounded condition wait
test_consumer.py stop-during-reclaim	reclaim_min_idle_ms=0: no clock
test_main.py e2e: two polls, 100 × 0.1 s	Bounded condition waits. The test takes about 2.6 s against 10 s per wait (about 4×)
test_child.py test_timeout_kills_the_child: timeout_s=2, < 10	5×
test_child.py terminal interrupt: 5 s pid poll, 1.5 s child sleep	A bounded wait for startup. SIGINT lands within one 50 ms poll of the pid write, against 1.5 s (30×)
test_contain.py undumpable: 10 s ready poll	A bounded condition wait
test_checkin.py test_a_stall_is_cut_off: 0.5 s stall against a 0.1 s timeout	5×, and load only lengthens the stall
test_deploy.py CI wait: < 15 s against a 30 s poll	The test takes 0.18 s

Verification

• ruff check, ruff format --check: clean. Full suite: 554 passed, 4 skipped.
• On co-processor, under 4 CPU-hog processes on 2 vCPU: the strand and deadline tests 50/50, then tests/test_consumer.py 50/50. The hogs' 900 s timeout ran out near the end of that second loop, so its last few runs had no load.

Review (CR --fix)

Item	Action	Result
1	Fixed	test_drift.py: the clock now reads 9.875, so the call has 0.125 s left, and the test matches read timeout=0.125). match="Timeout" had let a left / 2 mutation pass; it fails now
2	Fixed	test_main.py e2e: both waits are now bounded at 30 s by asyncio.timeout, and a processor run that exits is named at once. A failed start used to fail the test 21 s later with an unpacking error, then a ProcessLookupError from the SIGTERM; now it fails in 1.7 s with "processor run exited before its loop was up"

Gate after CR 2: ruff clean, uv lock --locked ok, full suite 554 passed / 4 skipped, scratch Redis clean. Both touched tests passed 20/20 in a loop.

🤖 Generated with Claude Code