A frozen clock for tests: one Agent holding one instant.

It ships in lib/ rather than test/support/ because host test suites need it, exactly as AuroraMeter.Test does. Drive it through AuroraMeter.Test.with_clock/2 and AuroraMeter.Test.travel/1,2 rather than starting it by hand: the helper serialises against other configuration changes, restores the previous clock: value on the way out (including on a raise) and stops the agent.

It is a single named agent, so it is global to the node: every test that installs it must be async: false.

All four readings come from the one frozen instant, db_now/0 included. That is the point of putting db_now/0 on the behaviour rather than leaving it as a bare query helper: a decision that consults the database's clock is exactly as testable as one that consults the node's, and a test can drive a cooldown or a retry horizon across its boundary in both directions without waiting.

monotonic_ms/0 moves in step with the frozen instant: travelling forward by 90 seconds moves it by exactly 90_000. It is an offset from the real monotonic reading taken when the agent started, not the frozen instant's epoch milliseconds, so it shares a number line with the real monotonic clock and a span measured across the end of a frozen block is still comparable with one measured outside it. A cache TTL, for example, does not come back with an expiry decades in the future.

Because db_now/0 is answered from the frozen instant, a test using this clock stops exercising the real query. AuroraMeter.ClockTest closes that gap deliberately with cases that call AuroraMeter.Clock.System.db_now/0.