Motivation

• Reduce steady-state cluster-bus bandwidth by using a lightweight PING/PONG header for routine failure detection while preserving existing safety semantics.
• Avoid regressions where topology/metadata and failover-sensitive flags (e.g. NOFAILOVER) become slow to propagate under light heartbeats.
• Ensure recovered-node failure-report cleanup remains reliable by retransmitting clear signals until stale reports can no longer affect quorum decisions.

Description

• Add light-heartbeat support guarded by a new capability bit CLUSTER_NODE_LIGHT_HDR_PING_SUPPORTED and per-node bookkeeping fields fail_report_clear_time and last_full_heartbeat_sent in clusterNode (see src/cluster_legacy.h).
• Split gossip processing into clusterProcessGossipEntries() to allow both full clusterMsg and light clusterMsgLight handling, and add clusterMsgLight count helpers to validate/process light packets (src/cluster_legacy.c).
• Implement clusterSendPingWithOptions() and clusterShouldSendFullHeartbeat() to choose between light and full heartbeats; light heartbeats carry compact failure gossip and retransmitted clear entries while full heartbeats still include topology, extensions and metadata (src/cluster_legacy.c).
• Preserve fast propagation for header-only metadata by forcing immediate full broadcasts from clusterUpdateMyself*/update helpers (use CLUSTER_TODO_BROADCAST_ALL) and ensure periodic per-peer full heartbeats at the configured direct-ping interval (cluster_ping_interval or cluster_node_timeout/2).
• Keep recovery-clear behavior deterministic by retransmitting recovered-node "clear" gossip until CLUSTER_FAIL_REPORT_VALIDITY_MULT * cluster_node_timeout expires, preventing stale external reports from causing false FAIL decisions (src/cluster_legacy.c).
• Add unit test coverage and small test adjustments: extend tests/unit/cluster/failure-marking.tcl with a retransmit/regression scenario and adjust tests/unit/cluster/cross-version-cluster.tcl to reflect mixed-version expectations.

Testing

• Built the server successfully with make -C src valkey-server -j4 (binary built).
• Ran ./runtest --single unit/cluster/failure-marking --clients 1 and the suite completed with all tests passing (regressions for recovery-clear addressed by the implementation).
• Ran a set of related cluster unit tests: unit/cluster/no-failover-option, unit/cluster/hostnames, unit/cluster/announce-client-ip, unit/cluster/announce-client-ports, and unit/cluster/announced-endpoints, and all executed tests passed.
• Ran ./runtest --single unit/cluster/cross-version-cluster --clients 1; mixed-version scenarios that require an external older binary were skipped in this environment (these tests are preserved and exercise fallback/compatibility when run with --other-server-path).

Codex Task