fix(swarm): stop treating silent idle agents as heartbeat failures

[Tong-bit-art]

Problem and change

UnifiedSwarmCoordinator.checkHeartbeats() marks every agent error once now - agent.lastHeartbeat > heartbeatIntervalMs * 3 (15s with the default 5s interval). Idle agents have no heartbeat producer and no in-flight work, so after registration they receive no messages and never refresh lastHeartbeat. Healthy idle workers therefore:

• flip to error after ~15s;
• disappear from getAvailableAgents() (which only returns status === 'idle'), so generic task dispatch stops using them;
• remain there until autoRecovery decays health to 0.2 (~75s) and resets them, after which the cycle repeats.

The coordinator only treats terminated as a non-live state. This change skips idle agents that also have no currentTask:

if (agent.status === 'terminated' || (agent.status === 'idle' && !agent.currentTask)) {
  continue;
}

Busy agents keep the existing liveness timeout: an agent that is actually working must still report a heartbeat or completion. The behavior for error/completed/other states is unchanged.

Validation

• 2 new regressions in v3/@claude-flow/swarm/__tests__/idle-agent-heartbeat.test.ts:
  • a silent idle agent stays idle with health 1 after a simulated timeout;
  • a silent busy agent is still marked error.
• Against main 19fb5c6e7: 1/2 fail (the idle case becomes error; the busy control passes).
• Candidate: 2/2 pass; full @claude-flow/swarm suite 346/346, 31/31 files; tsc --noEmit 0 errors; git diff --check clean.

Scope / known limitations

• This does not add a heartbeat producer for real worker processes. A truly dead idle agent is only discovered when it is assigned work; before this change it was reported dead while healthy, which was a false positive.
• AgentRegistry (the deprecated SwarmHub path) has its own heartbeat loop and is unchanged.
• Domain-pool membership/teardown and terminal-owner cleanup are covered by the already-merged fix(swarm): stop domain pools from creating unregistered agents #4246/fix(swarm): release domain-pool agents after terminal tasks #4247/fix(swarm): remove domain-pool membership when an agent is unregistered #4269/fix(swarm): release stale task owners on terminal states #4272 changes.
• No public API or return-shape changes.

Compatibility

• Only a private liveness check changes; no method signatures, events, persistence formats, dependencies, or lockfiles.
• Agents with an active task still use the same 3-interval timeout and auto-recovery path.

[ruvnet]

Verified the mechanism directly rather than taking the write-up at face value: the old `checkHeartbeats()` loop flagged any non-terminated agent as `error` purely on heartbeat silence, with no distinction for agents that have no heartbeat producer to begin with (idle, no current task). The fix's skip condition (`agent.status === 'terminated' || (agent.status === 'idle' && !agent.currentTask)`) is correctly scoped — it only exempts truly-idle agents, and a busy agent that goes silent still hits the existing 3-interval timeout path unchanged (confirmed in the diff: the busy-agent branch is untouched below the new early-`continue`).

The two new tests precisely isolate the two cases (idle+silent stays healthy; busy+silent still flips to `error`), which is exactly what's needed to pin this behavior down. Small, surgical, low blast radius — no public API/signature changes, one private liveness check.

Will merge once CI is green.