Fix live regression: NativeAck endpoints get an unregistered BatchedSender

[jeremydmiller]

Live regression on main, introduced by #4061 (mine). Sending anything to a NativeAck endpoint on Redis fails outright:

System.InvalidOperationException: This sender has not been registered.

Mechanism

1. GH-3709: send inline from NativeAck endpoints to close the interceptor loss window #4061 remapped EndpointMode.NativeAck from BufferedSendingAgent to InlineSendingAgent, to close the global-partitioned interceptor's loss window.
2. InlineSendingAgent is ISendingAgent, IDisposable — not an ISenderCallback. So the registration in EndpointCollection.CreateSendingAgent is skipped:

   if (sender is ISenderRequiresCallback senderRequiringCallback && agent is ISenderCallback callbackAgent)
   {
       senderRequiringCallback.RegisterCallback(callbackAgent);
   }

3. Every transport that uses BatchedSender selects it by asking Mode == EndpointMode.Inline. NativeAck is not Inline, so those transports still build one.
4. BatchedSender throws whenever _callback is null — on send, on ping, and on the failure path.

Redis adopted NativeAck in #4056, so this is broken on main right now.

Why no PR's CI caught it

RabbitMQ and Pulsar use their own senders rather than BatchedSender, and they are the only transports with NativeAck integration tests. Redis has such tests, from #4056 — but #4061 branched from main before #4056 merged, so #4061's CI ran against a tree that did not contain them.

Both PRs were green. The combination is red. This is the classic stale-base gap, and it is worth knowing that nothing in the current workflow would have caught it: neither branch was wrong on its own.

The fix

One property rather than nine copies of a boolean:

public bool SendsInline => Mode is EndpointMode.Inline or EndpointMode.NativeAck;

Applied to the sender gates in SNS, SQS, Azure Service Bus (topic and queue), Kafka, MQTT and Redis.

It is centralised deliberately. The failure mode is silent and total — an endpoint that looks configured and cannot send a single message — and the per-transport form invites exactly this bug again the next time a mode is added. The xml-docs on SendsInline state the coupling to ISenderCallback explicitly, so the next person changing a sending-agent mapping can see why the property exists.

GCP Pub/Sub is deliberately untouched. It has not adopted NativeAck (supportsNativeAck is false), so it cannot reach this today, and that file is being actively edited under #4065/#4066. Its gate should move to SendsInline when #4052 adopts the mode.

Verification

• Red-baselined: with the Redis gate reverted to Mode == EndpointMode.Inline, three of six Wolverine.Redis.Tests.native_ack_mode tests fail with the exact "This sender has not been registered." exception. With SendsInline, 6/6 pass.
• dotnet build wolverine.slnx -c Release -f net9.0 — clean, 0 warnings, 0 errors
• CoreTests — 2581 total, 0 failed, 2 skipped
• Wolverine.Redis.Tests.native_ack_mode — 6/6

Credit

Found by the agent working #4049 while investigating what looked like emulator contention. A 10-second assertion failure is not a timeout, so it traced the cause instead of writing it off — which is the only reason this was caught before release rather than by a user.

🤖 Generated with Claude Code