Performance: reduce per-call allocations in channel recreation outcome tracking

[berndverst]

1. What is the issue?

When channel recreation is enabled, every unary gRPC call registers a continuation and boxes a tuple state object solely to update consecutive-failure tracking.

2. Likely priority

Low. This is a small cost per call, but it can become observable for high-QPS instance-management or signaling clients.

3. Impact on performance

Each unary call allocates continuation-related state and a boxed (ChannelRecreatingCallInvoker, methodName) tuple. The resulting continuous allocation rate can increase GC pressure even when all RPCs succeed.

4. Details and code reference

Unary-call interception:

durabletask-dotnet/src/Client/Grpc/ChannelRecreatingCallInvoker.cs

Lines 96 to 103 in 883211a

	public override AsyncUnaryCall<TResponse> AsyncUnaryCall<TRequest, TResponse>(
	Method<TRequest, TResponse> method, string? host, CallOptions options, TRequest request)
	{
	TransportState current = Volatile.Read(ref this.state);
	AsyncUnaryCall<TResponse> call = current.Invoker.AsyncUnaryCall(method, host, options, request);
	this.ObserveOutcome(call.ResponseAsync, method.FullName);
	return call;
	}

Outcome observation:

durabletask-dotnet/src/Client/Grpc/ChannelRecreatingCallInvoker.cs

Lines 198 to 218 in 883211a

	void ObserveOutcome<TResponse>(Task<TResponse> responseAsync, string methodFullName)
	{
	// Use ContinueWith with TaskScheduler.Default so we don't capture sync context.
	responseAsync.ContinueWith(
	(t, state) =>
	{
	var (self, name) = ((ChannelRecreatingCallInvoker, string))state!;
	if (t.Status == TaskStatus.RanToCompletion)
	{
	self.RecordSuccess();
	}
	else if (t.Exception?.InnerException is RpcException rpcEx)
	{
	self.RecordFailure(rpcEx.StatusCode, name);
	}
	},
	(this, methodFullName),
	CancellationToken.None,
	TaskContinuationOptions.ExecuteSynchronously,
	TaskScheduler.Default);
	}

ObserveOutcome registers ContinueWith and passes (this, methodFullName) through the object-typed continuation state parameter.

5. Guidance for how to fix

Profile allocation rates first. If material, redesign the observation helper to avoid boxed state while preserving current success and transport-failure classification, execution-synchronous continuation behavior, and method-name diagnostics. Do not compromise the existing lock-free channel lifecycle or failure accounting for a micro-optimization.