Dial-path disconnect no longer waits for the peer's Close frame

[chrisuthe]

Closes #160. Independent of #167 — no file overlap.

Both reasons for deferring this were wrong

#160 was filed on shape rather than evidence, and gave two reasons its risk profile differed from #143. Checking them is what this PR started as, and neither holds:

it is followed by CleanupWebSocketAsync in a finally, so the socket is released regardless

A finally runs when the try exits. An await that never returns never exits, so CleanupWebSocketAsync is unreachable in exactly the case it looks like it covers.

it is not on a disposal path invoked with CancellationToken.None

DisposeAsync (SendspinConnection.cs:622) calls await DisconnectAsync(GoodbyeReasons.Shutdown) with no token, so cancellationToken is default — CancellationToken.None. That is the disposal path with None.

It is worse than "no caller happens to pass a token": ISendSpinClient.DisconnectAsync(string reason) has no CancellationToken parameter, so no caller can pass one. Every call site in src/ passes only a reason.

So the dial path sits in the same position the listen path did before #143 — an initiating close awaiting a frame the peer may never send, on the disposal path, with a token that cannot fire. An app exit against an unresponsive server hangs indefinitely.

The fix

CloseOutputAsync instead of CloseAsync, matching WebSocketClientConnection and carrying the reasoning at the call site — including why the finally and the token do not save it, since that is the exact inference that deferred this once already.

Verification

The peer completes the RFC 6455 upgrade and then never reads the socket again, so it cannot answer a Close frame. That makes this a statement about the shape of the call rather than a race: the outcome is the same on every run.

	Result
CloseOutputAsync (this PR)	passes in 61 ms
CloseAsync (reverted)	fails at the 5 s bound

The test bounds its own wait with Task.WhenAny rather than letting a regression hang. A hanging test burns the whole CI job instead of failing — which is how #143 stayed invisible — and the peer is dropped in a finally before the assert so a parked close faults rather than wedging the run.

• Clean dotnet build -c Release --no-incremental: 0 errors, both library TFMs.
• Full suite: 730 passed, 0 failed (729 + 1 — count checked).

Release notes updated: the entry now covers both teardown paths and names DisposeAsync as the reason it mattered.