GH-3520: EventSubscriptionAgent restores continuous execution after Rebuild/Rewind

[jeremydmiller]

Fixes #3520. Wolverine-side half of a coordinated fix with JasperFx/jasperfx#536.

The bug

Under Wolverine-managed event-subscription distribution there is no store coordinator (AddAsyncDaemon) to resurrect a shard the daemon stopped:

• EventSubscriptionAgent.RebuildAsync → _daemon.RebuildProjectionAsync(...) stops the continuous agent and never restarts it. The wrapper still reports Running against the now-stopped agent, so NodeAgentController sees nothing to fix.
• EventSubscriptionAgent.RewindAsync → _daemon.RewindSubscriptionAsync(...) restarts the agent daemon-side, but the wrapper's _innerAgent still points at the stopped pre-rewind agent and Status still reads Running.

Either way the shard re-registers RegisteredIdle daemon-side and freezes forever while the high-water mark climbs (JasperFx/CritterWatch#747). An operator "Restart" routed to an agent Wolverine believes is already running no-ops.

The fix

RebuildAsync/RewindAsync now restore continuous execution themselves through the registered daemon start path (resumeContinuousAsync) and refresh _innerAgent/Status:

private async Task resumeContinuousAsync(CancellationToken cancellationToken)
{
    _innerAgent = await _daemon.StartAgentAsync(_shardName, cancellationToken);
    Status = AgentStatus.Running;
}

resumeContinuousAsync deliberately does not re-invoke OnStarted: a rebuild/rewind is transparent to EventStoreAgents' per-node running-agent count (the wrapper never observed a matching stop), so re-counting would leak the database's tracker observer subscriptions at the 1→0 transition.

Coordination / merge order

• Rebuild half is independently correct. RebuildProjectionAsync leaves the shard stopped-but-registered, so _daemon.StartAgentAsync(_shardName) cleanly starts and re-registers a fresh continuous agent.
• Rewind half depends on In Memory Queues don't seem to work when IDocumentSession is injected into EndPoint's method #536. Before that fix, RewindSubscriptionAsync leaves an unregistered running agent, so re-calling StartAgentAsync would start a second agent on the same progression row. In Memory Queues don't seem to work when IDocumentSession is injected into EndPoint's method #536 registers the rewind's restarted agent, making the wrapper's start resolve to that same running agent idempotently. Merge in lockstep with the JasperFx.Events bump that includes In Memory Queues don't seem to work when IDocumentSession is injected into EndPoint's method #536.

Tests

event_subscription_agent_rebuild_rewind_resume (CoreTests) drives RebuildAsync/RewindAsync over a substituted IProjectionDaemon and asserts each resumes through the registered start path (StartAgentAsync received once) and reports Running. Both verified red before the fix, green after.

🤖 Generated with Claude Code

https://claude.ai/code/session_018LDiv9GqbQkkAuU1nf4H6S

---

Related issues

• Fixes EventSubscriptionAgent.RebuildAsync/RewindAsync leave the shard permanently stopped under Wolverine-managed distribution (daemon stops continuous agents; nothing restarts them) #3520
• Coordinated with (depends on) Daemon start-registration + failure-reason fixes (closes #534, #535; wolverine#3519/#3520) jasperfx#536, which closes JasperFxAsyncDaemon.StartAgentAsync throws a causeless 'Unable to start a subscription agent' — failed-start reasons are swallowed (startContinuousShardAsync bool ignored, silent false paths) jasperfx#534 and Daemon rebuild/rewind lifecycle gaps on coordinator-less hosts: rebuildProjection never restarts the continuous agents it stops; RewindSubscriptionAsync restarts agents that bypass the _agents registry jasperfx#535
• Companion Wolverine issue: Managed event-subscription distribution: randomized agent fails first start on multi-(ancillary-)store Marten hosts and wedges in a permanent 30s retry loop #3519