Issue #3230 changes only the SQLite restart test fixture. The real Agent.Service loop remains the implementation, wrapped solely with Stream.provideService(Clock.Clock, controlledWallClock). Host control/engine owners, polling and journal followers retain live Clock; controlled clock sleep and monotonic time delegate to that live Clock. Budget policy, admission, suspend/resume recovery and production source stay unchanged. No ceiling is relaxed: 500ms original, 1600ms exact raise; 1100ms spend and 1500ms parked span are durably recorded. Completing 100ms resumed work is allowed, while 600ms resumed work re-parks with 1700ms spend before a third provider call. Existing Stop, unasked replay, token, USD, model-call, cell and tool-call controls are retained. The helper waits for an approval to become an actual durable budget wait before the test composition closes, avoiding teardown release racing the park. Historical setup errors (missing prototype sleep in initial Clock spread, broad host clock timeout, and teardown released/budget race) are retained as setup diagnostics; none is reported as the original product failure. Original failure is the issue reporter's 3-of-3 loaded-host evidence; raw original subprocess logs are unavailable to this subagent. No production coverage percentage or whole-package 100% claim follows from the focused, coverage-disabled test command.
