[mgravell]

Addresses #3284 (cross-slot commands refused client-side on a non-clustered Azure Managed Redis after a connection blip on 3.4.5).

Diagnosis

Reproduced with the in-process server (HelloClusterModeUnitTests): a server whose HELLO says mode: cluster while INFO says redis_mode:standalone, with every CLUSTER subcommand refused.

• The subscription connection's handshake is HELLO alone (no INFO, no CLUSTER), and under RESP3 pub/sub has had its own connection since Use a dedicated pub/sub connection under RESP3 by default #3264. AMR prefers RESP3.
• When only that leg reconnects (the /Subscription SocketClosed in the report), its HELLO retypes the server as Cluster and nothing on that leg corrects it. On first connect the interactive leg's INFO lands last and wins, which is why it ran for 24h first.
• The type change triggers a reconfigure; with an already-connected interactive leg that is a PING tracer only, so the server now counts as a cluster and the multiplexer enters cluster mode with no slot map. Cross-slot commands then throw. This is also the path that produced RedisServerException "ERR command is not allowed" after v3.4.0 update #3274 (same reporter, pub/sub via SignalR).

Still worth measuring on a real Enterprise-policy AMR database: HELLO 3 mode versus INFO server redis_mode. The fix holds either way, but it decides whether the refused-CLUSTER test model should say cluster in INFO or only in HELLO.

Changes

• Cluster routing requires a slot map. Servers that identify as a cluster but give us no map are routed as a single standalone server (new event id 133). The map is never cleared once populated, so an established cluster whose refresh is refused (Tolerate refused CLUSTER SLOTS/NODES during topology updates #3275) keeps routing as one. When a map arrives, stale redundant-primary flags from an earlier election are cleared.
• AnyServer accepts a cluster-typed server while routing as standalone (ServerType.IsRoutableAs); it previously required exact equality, which dropped every connection in that mode.
• HELLO mode/role are applied only from the interactive connection. Per-server facts come from the connection that does discovery; the subscription leg must not overwrite them.
• Toy server: overridable HelloModeValue so HELLO and INFO can disagree. Docs: one sentence in Configuration.md.

Deliberately not done: HELLO version/vendor on the subscription leg are still applied (harmless, both legs agree); could be gated the same way for consistency.

Verification

• New tests: pass under RESP2 and RESP3 (RESP3 fails before the change with the exact exception from the report).
• ClusterCommandsRefusedUnitTests extended to assert standalone routing and a working cross-slot MGET.
• All in-process-server test classes, net10.0: 645 passed, 16 skipped.
• Full suite against the docker topology, twice: 6661 passed, 161 skipped, 1 failed. The failure is FailoverTests.SubscriptionsSurvivePrimarySwitchAsync (RESP2), which fails at the same rate (2 of 3 class-level runs) on a baseline worktree at the branch point and passes in isolation: pre-existing flake, unrelated.
• dotnet build Build.csproj -c Release /p:CI=true: 0 warnings, 0 errors.

[mgravell]

Follow-up commit (8d6e3cc): the routing decision now uses ServerSelectionStrategy.IsSlotMapEmpty (derived from whether the map was allocated, O(1)) instead of the scan; CountCoveredSlots stays only for the log line in the cluster branch, as before. UpdateClusterRange ignores inverted or out-of-range ranges so the map is never allocated without an owner being written (previously an out-of-range slot would have thrown IndexOutOfRange there). SlotMapUnitTests covers the rejected ranges. Release build clean; in-process test classes 646 passed.