GH-4186: report a NativeAck listener's real queue depth and last receipt

[jeremydmiller]

Closes #4186.

What was wrong

ListeningAgent read the receiver's depth through ILocalQueue:

public int QueueCount => (_receiver is ILocalQueue q ? q.QueueCount : 0) + ...

NativeAckReceiver and InlineReceiver are deliberately not local queues — a delivery in either mode settles against the listener that brought it, not against a queue — so both contributed a constant 0 to EndpointHealthSnapshot, despite each already maintaining a real depth over a real block. The issue's inference was correct: a saturated NativeAck listener was indistinguishable from an idle one.

LastQueueActivityAt was the same root cause. Its change-detection heuristic has exactly one writer, UpdateQueueCountObservation(), whose only caller is BackPressureAgent — which correctly does not run for these two modes (Endpoint.ShouldEnforceBackPressure, GH-3708). So the timestamp stayed at the DateTimeOffset.UtcNow it was initialised with at construction, for the life of the process.

No back-pressure behaviour was changed for NativeAck or Inline. They still have no BackPressureAgent, as designed.

The shape

Shape 2 from the issue, and the test pins why rather than just the outcome. A new narrow IHasQueueDepth that ILocalQueue extends:

• BufferedReceiver / DurableReceiver satisfy it for free, unchanged.
• NativeAckReceiver / InlineReceiver implement it directly — reporting a depth without claiming they can be enqueued into.

Shape 1 (NativeAckReceiver : ILocalQueue) was the smaller diff and the wrong one: EnqueueDirectlyAsync type-switches on ILocalQueue before it reaches the NativeAck branch GH-4011 added, so a NativeAck receiver claiming to be a local queue would silently take the wrong path on every DLQ replay. a_native_ack_receiver_reports_a_depth_without_claiming_to_be_a_local_queue guards that.

LastReceivedAt is a default-null member on the same interface, stamped on receipt by the two receivers that need it. ListeningAgent prefers it when it is the more recent of the two, so every other receiver keeps the existing heuristic byte-for-byte.

One thing beyond the issue, called out deliberately

The two pass-through wrappers — ReceiverWithRules and GlobalPartitionedInterceptor — were also swallowing the depth. Delegating through them is required for the NativeAck fix to hold when incoming envelope rules or a global-partitioned topology are in play.

For ReceiverWithRules that is reporting only. For GlobalPartitionedInterceptor wrapping a Buffered or Durable receiver it is not: QueueCount read 0, so BackPressureAgent could never fire on a global-partitioned endpoint and the buffer could grow unbounded. That is a real fix, but it is a behaviour change on a path the issue did not name, so flagging it rather than burying it.

Evidence

native_ack_queue_depth_4186 asserts on CollectEndpointHealth() end to end, exactly as the issue asked for — listener started through IEndpointCollection, handlers parked on a gate, snapshot read from the real collection.

Verified red first: with ListeningAgent reverted to the ILocalQueue read, depth_and_receipt_activity_reach_the_endpoint_health_snapshot times out on QueueCount > 0 and inline_receiver_depth_reaches_the_snapshot_too gets 0 where it wants 1.

Green after: 3/3 here, and CoreTests 2674/2674 with 2 pre-existing skips. dotnet build wolverine.slnx -c Release -f net9.0 clean.

🤖 Generated with Claude Code