InternalFailure: "Output pipe not initialized" when a write races connection teardown

[mgravell]

Summary

A write that reaches PhysicalBridge.WriteMessageToServerInsideWriteLock while the underlying
connection is being torn down and re-established fails with an InvalidOperationException
("Output pipe not initialized"), surfaced to the caller as
RedisConnectionException: InternalFailure. Because it is an InternalFailure rather than a
recognised connection failure, it is not subject to the usual backlog/retry handling.

Seen intermittently; @mgravell reports having hit it before, so this is not a one-off.

Trace

StackExchange.Redis.RedisConnectionException : InternalFailure on [0]:UNWATCH (TransactionProcessor)
---- System.InvalidOperationException : Output pipe not initialized

----- Inner Stack Trace -----
   at StackExchange.Redis.PhysicalConnection.<get_Output>g__Throw|172_0() in /_/src/StackExchange.Redis/PhysicalConnection.Write.cs:line 18
   at StackExchange.Redis.PhysicalConnection.get_Output() in /_/src/StackExchange.Redis/PhysicalConnection.Write.cs:line 17
   at StackExchange.Redis.Message.WriteTo(PhysicalConnection physical) in /_/src/StackExchange.Redis/Message.cs:line 950
   at StackExchange.Redis.PhysicalBridge.WriteMessageToServerInsideWriteLock(PhysicalConnection connection, Message message) in /_/src/StackExchange.Redis/PhysicalBridge.cs:line 1612

What provoked it

Observed during a full test-suite run, in TransactionTests.BasicTranWithStreamLengthCondition.
The relevant detail is what was happening on the shared connection at that moment: a parallel
ClientKillTests.ClientKillWithMaxAge had just issued a CLIENT KILL ... MAXAGE, killing the
interactive connection, and the bridge was mid-backlog-drain when the reconnect happened.

From the same test's output, immediately before the failure:

Interactive/127.0.0.1:6379: MessagesSinceLastHeartbeat: 317, ConnectedAt: 2026-08-06 14:30:14Z,
  Writer: Active, BacklogStatus: WritingMessage,
  BacklogMessagesPending: (Queue: 193, Counter: 193), TotalBacklogMessagesQueued: 369,
  Connection: (SentAwaitingResponse: 0, AvailableOnSocket: 0 byte(s), InReadPipe: 0 byte(s),
               InWritePipe: -1 byte(s), ReadStatus: ReadAsync, WriteStatus: Writing)

and from the parallel test that did the killing:

Connect took: 3025ms
Connection Failed (Interactive,SocketClosed): ... SocketClosed on 127.0.0.1:6379/Interactive,
  Idle/RanToCompletion, last: CLIENT, origin: ReadAllAsync, ... v: 3.1.12.21783
Connection Restored (Interactive,None):

So: connection killed out from under an active backlog drain, and a queued message is then written
to a PhysicalConnection whose output pipe has already gone.

Expected

A write losing its connection mid-flight is an ordinary connection failure, and should be handled
as one (backlog per BacklogPolicy, or a RedisConnectionException describing the disconnect),
not an InternalFailure carrying an InvalidOperationException.

Environment

• main @ 5517cbf2
• net10.0 (.NET 10.0.10), library reporting v: 3.1.12.21783
• Linux, server = redislabs/client-libs-test image, Redis 8.9.241
• Reproduction is not reliable; killing connections while a backlog is draining looks like the way in.