[mgravell]

Adds first-class support for CLUSTER SLOT-STATS (Redis 8.2): per-slot usage statistics, so callers can see which slots carry a cluster's load. Promotes the deferred row in CommandCoverage ("pairs with hot-key tooling" — the hot-key tooling has since shipped).

API (all experimental, SER012)

// the two server forms, as two IServer entry points (strict oneof grammar; overloads would trip RS0026)
ClusterSlotStatsResult? ClusterSlotStatsBySlotRange(int startSlot, int endSlot, ...);          // SLOTSRANGE start end
ClusterSlotStatsResult? ClusterSlotStatsOrderBy(ClusterSlotStatsMetric metric,
    long limit = 0, Order order = Order.Descending, ...);                                      // ORDERBY metric [LIMIT n] [ASC]

ClusterSlotStatsEntry carries Slot, KeyCount (always collected), and CpuTime (TimeSpan?, from cpu-usec), MemoryBytes, NetworkBytesIn/Out as nullable values: the server only reports what cluster-slot-stats-enabled collects, and "not collected" is deliberately distinct from a collected zero. Metrics the library doesn't recognize are preserved in Metadata (the metric set is documented as extensible, and has grown before).

Design grounding

• File layout, message, nil handling: traced from CLUSTER SLOTS (ClusterSlots.cs siblings) — partial-interface file per feature, Message.Create(-1, flags.WithCategory(NodeLocalRead), RedisCommand.CLUSTER, ...), lenient aggregate parsing, null for nil.
• Reply-field handling: traced from HOTKEYS — internal [AsciiHash] field enum with generated TryParse, nominal (never positional) parsing, microsecond counters surfaced as TimeSpan, limit = 0 meaning "server default" as HotKeysStart(count = 0) does.
• Minimal wire form: LIMIT sent only when limit > 0, ASC only when ascending — the server defaults are expressed by omission.
• Versioning: RedisFeatures.ClusterSlotStats gates on the existing v8_2_0_rc1 constant. Pre-8.2 answers ERR Unknown subcommand.
• SER012 is a feature-specific experiment id (SER011 is Server 8.12 on the open Set-member expiration (Redis 8.12): SADDEX / SEXPIRE / SPTTL / SPERSIST on IDatabase #3340; keeping them separate avoids coupling the PRs).

Plan with observed server facts (verified against src/cluster_slot_stats.c and a live cluster): notes/cluster-slot-stats/plan.md.

Tests

• Round-trip (RoundTripUnitTests/ClusterSlotStatsRoundTrip.cs, 10/10): exact request bytes for both forms including LIMIT/ASC minimality; the full metric map parsed identically from RESP2 flat pairs and a RESP3 map; key-count-only replies leave the gated metrics null; an unknown metric lands in Metadata; empty array and nil replies; argument validation.
• Live (ClusterSlotStatsTests.cs, 12/12 against the repo cluster topology, per protocol, zero skips): a routed write is counted in its slot by the owning node; whole-range answers stay shard-local and ascending; ORDERBY honours LIMIT and both orders; CONFIG SET cluster-slot-stats-enabled "cpu net" surfaces CpuTime/NetworkBytes* at runtime (with MemoryBytes still null, config restored in finally); ORDERBY on the disabled mem metric gets the server's ORDERBY refusal; the standalone node answers "cluster support disabled".

Full solution build clean (/p:CI=true, 0 warnings) across net461/net472/netstandard2.0/net8.0/net10.0; eng/public-api.py --check passes.

Note: RS0016 does not fire in my local environment, so the PublicAPI.Unshipped.txt entries are hand-written against the shipped [SER009] / #3340 [SER011] precedents and sorted with eng/public-api.py — worth a glance.