[jeremydmiller]

Refs JasperFx/wolverine#5318

JasperFxAsyncDaemon.StopAllAsync calls HighWaterAgent.StopAsync, which awaits the poll loop. The loop's wait (IDaemonWakeup.WaitAsync with FastPollingTime or SlowPollingTime) was only tied to the daemon's own token, which a stop does not cancel, so every stop waited out the rest of the current tick: about 250 ms after activity, up to 1 s after a quiet spell. Wolverine's PauseThenCatchUpOnMartenDaemonActivity pays that on every tracked call.

Change

• Each poll loop gets its own CancellationTokenSource, linked to the daemon token and created in launchAsync, so every start (and every watchdog restart) has a fresh one and stop-then-start keeps working.
• The loop's waits use that token. StopAsync cancels it before awaiting the loop; the cancelled wait is the normal exit path (waitAsync already swallows it, and the loop then sees IsRunning == false and leaves).
• A restart cancels and disposes the retired loop's source, so an abandoned loop parked in a wait exits immediately instead of on its next tick. The generation guard is unchanged.
• Detect calls still use the daemon token; only the waits changed.

I looked at the other daemon loops. Subscription agents are fed by the high-water mark rather than polling, and the coordinator loops in ProjectionCoordinatorBase already wait on their stopping token, so nothing else has this shape.

Tests

New HighWaterAgent_stop_cancels_poll_wait_Tests: FastPollingTime / SlowPollingTime = 5 s, stop must return in under 1 s, then start again, check the new loop polls, and stop again in under 1 s.

• Without the fix: fails, stop took 4.9 s.
• With the fix: passes (488 ms total).
• EventTests --filter-namespace "EventTests.Daemon*": 499/499 on net9.0 and 499/499 on net10.0.

🤖 Generated with Claude Code

https://claude.ai/code/session_01DCggXgxnG3EBXoR9rrq37U