This is the promise the whole project was built to keep: a real process supervisor. The script exercises every layer at once. Dependency ordering holds web down until db is genuinely Running, not merely spawned. The restart policy and backoff bring flaky back after each crash - at t=1, t=3, then t=7, the intervals 1, 2, 4 seconds apart exactly as the doubling predicts. And crash-loop protection ends the madness: after three restarts inside the window, the fourth crash does not schedule a fifth attempt - flaky is given up as Failed, permanently. The exact final Status is only reachable if the state machine, the policies, the clock, the backoff, the window, and the ordering all agree.

Then the shutdown closes the loop: SIGTERM to the running services in reverse dependency order, web before db, flaky skipped because a Failed service has nothing to stop, and both reaped cleanly to Stopped. From a bare service record that only knew its own name, you have built the honest core of a real supervisor - a six-state lifecycle over an injectable runtime, four restart policies, exponential backoff with crash-loop give-up, a dependency graph with topological start and reverse-order shutdown, and a reconcile loop - the same design that sits inside runit, supervisord, and systemd, minus the operating-system layer that fork/execs real processes and delivers real signals. That layer is the runnable entry point this reference adds on top; the engine underneath is yours, and it is exact.