#816 Production-readiness: fault-injection / chaos harness for the durable engine (#agentic)

Add tests/chaos: a fault-injection harness that SIGKILLs workers mid-flight and asserts durability invariants (I1 no lost/stuck, I2 exactly-once side effects, I3 no #810 corruption). Canary exercises inline task+checkpoint, durable sleep/wake, activity-worker routing, and a transactional exactly-once marker (tools.chaos.mark -> chaos_marks). This is the first real production-readiness gate beyond the golden harness.