[mgravell]

Sub-PR of #3343. IDatabase support for the sorted-set moves that Redis 8.12-rc3 adds (redis/redis#16135): ZMOVE and ZMOVEMEMBER. Written against the client HLD in redis-developer/client-designs (redis-oss/client-hld/zmove/README.md). BZMOVE is not exposed, per the blocking-command policy.

Surface

Member	Command	Notes
SortedSetMove(src, dst, Order order = Ascending, long? count = null)	ZMOVE src dst MIN|MAX [COUNT n]	from the low or high end
SortedSetMoveByScore(src, dst, double start, double stop, Exclude exclude = None, long? count = null)	ZMOVE src dst BYSCORE min max [COUNT n]	bounds as SortedSetRemoveRangeByScore takes them
SortedSetMoveMembers(src, dst, RedisValue[] members, When when = Always)	ZMOVEMEMBER src dst [NX|XX] MEMBERS n m...	When is the per-member condition on the destination

All return SortedSetEntry[]? (and Task<...> twins), and all are [Experimental(Experiments.Server_8_12)].

Choices worth a look:

• null means nothing moved. The server's null reply is never folded into an empty array; same contract as the LMOVEM overload of ListMove.
• COUNT passes through. It is sent if and only if the caller gives one, including 1 and values the server rejects. SortedSetPop elides COUNT 1; that is not copied, because for BYSCORE no count means the whole range.
• Three names rather than overloads. SortedSetMove(a, b, 0, 5) would be ambiguous between Order/long? and double/double (literal 0 converts to an enum).
• start/stop are required on SortedSetMoveByScore: infinite defaults would make a bare call move the whole set.
• Members are not tidied. Not de-duplicated or re-ordered, a member named NX or MEMBERS is data, and an empty list goes out as MEMBERS 0 for the server to reject.

Reply parsing

The reply is nested [member, score] pairs in RESP2 as well as RESP3 (only the score type differs), unlike ZPOPMIN. SortedSetWithScores reads RESP2 as a flat interleaved list, so there is a new strict processor, ResultProcessor.NullableSortedSetEntryArray; a flat list or a non-pair is an unexpected reply rather than a guess.

Routing, retry, coverage

• Both commands are primary-only and CommandRetryWriteAccumulating, alongside ZMPOP.
• BZMOVE gets a policy row in docs/CommandCoverage.md; no enum member, as with BZMPOP.
• Transaction analyzer: nothing mapped, with the reason left next to the SMOVE row. SortedSetRemove + SortedSetAdd writes the caller's score and adds unconditionally; ZMOVEMEMBER carries the source's score and moves nothing when the source lacks the member.

Shared 8.12 plumbing

• New RedisFeatures.v8_12_0_rc3 (8.11.242) and RedisFeatures.SortedSetMove: the existing v8_12_0 is rc2 (8.11.241), which does not have these commands.
• Test image moves to client-libs-test:8.12-rc3 (Dockerfile and compose). rc3 requires cluster-bus-port-protected-mode no for a plaintext cluster bus; the cluster configs on this branch already set it.
• docs/exp/SER011.md gains the feature bullet.

Tests

• ResultProcessorUnitTests/SortedSetMove: RESP2 and RESP3 pairs, single pair stays nested, infinite scores, order, both null forms, empty array, six malformed shapes.
• RoundTripUnitTests/SortedSetMoveRoundTrip: exact bytes for every token shape, plus argument validation.
• SortedSetMoveTests (live, [RunPerProtocol], require: v8_12_0_rc3): selection, count, null cases, same key, conditions, keyword-like members, server errors, transaction, key prefix, and two cluster tests (same-slot routing; cross-slot refused with nothing moved).
• KeyPrefixed* forwarding tests for all six members.

Run locally: 76/76 on net10.0 and net8.0 against throwaway 8.12-rc3 standalone and single-node cluster servers. Full suite on net10.0 against an older (8.9.241) topology: 6838 pass, 254 skip, 1 fail (FailoverTests.SubscriptionsSurvivePrimarySwitchAsync, which passed on one of two isolated reruns; not checked on the base branch). net481 not run.

dotnet build Build.csproj -c Release /p:CI=true is clean apart from the NU1903 ILCompiler advisory in toys/AotRig.

Known, not addressed here

On a cluster, cross-slot keys are refused by the library for the bare SortedSetMove(src, dst) and by the server (CROSSSLOT) for the other forms. The cause is in Message.cs: the two-key messages carrying two or more values do not combine the second key into the hash slot. It predates this branch and same-slot calls route correctly; the cluster test accepts either layer. To be handled separately.