Bug: Cron timer stops firing after config change (v2026.2.3-1)

Environment: macOS Sequoia (arm64), Node 25.5.0, OpenClaw 2026.2.3-1

Description

After changing the default model in config (from claude-opus-4-5 to anthropic-cloud/claude-opus-4-6), all cron jobs stopped executing. The scheduler initializes correctly but the timer callback (onTimer) never fires.

Steps to Reproduce

1. Have working cron jobs (isolated agentTurn, every 5min, model: local-llm/gemini-2.5-pro)
2. Change agents.defaults.model.primary from one model to another
3. Restart gateway (SIGUSR1 or cold restart)
4. Observe: cron jobs never execute again

Observed Behavior

• cron: started log entry shows correct jobs count and nextWakeAtMs
• cron.status returns enabled: true with valid nextWakeAtMs
• No timer tick ever fires — zero cron-related log entries after startup
• cron run {jobId} returns { ran: false, reason: "not-due" } even when nextRunAtMs is in the past
• nextRunAtMs in jobs.json keeps advancing forward (gets recomputed) without execution
• No errors in logs

What I Tried (all failed)

• SIGUSR1 restart
• Full cold restart (launchctl bootout + bootstrap)
• Deleted ALL jobs + runs/ directory + jobs.json.bak, recreated fresh
• Both sessionTarget: "isolated" (agentTurn) and "main" (systemEvent)
• Both wakeMode: "now" and "next-heartbeat"
• Increased cron.maxConcurrentRuns to 4
• Minimal test job: { schedule: { kind: "every", everyMs: 60000 }, payload: { kind: "agentTurn", message: "Reply TEST_OK", model: "local-llm/gemini-2.5-pro" } }

Suspected Root Cause

In onTimer():

await locked(state, async () => {
    await ensureLoaded(state, { forceReload: true });
    await runDueJobs(state);  // <-- never finds due jobs?
    await persist(state);
    armTimer(state);
});

ensureLoaded with forceReload: true calls recomputeNextRuns(state) which recalculates all nextRunAtMs values to the future (via computeNextRunAtMs(schedule, nowMs)). By the time runDueJobs checks now >= nextRunAtMs, the values have already been pushed forward.

However, this code path existed before and worked — so something else may be involved (possibly related to the config change invalidating internal state).

Config (relevant parts)

{
  "agents": {
    "defaults": {
      "model": {
        "primary": "anthropic-cloud/claude-opus-4-6",
        "fallbacks": ["anthropic-cloud/claude-opus-4-5-20251101", ...]
      },
      "heartbeat": { "every": "1h", "model": "anthropic-cloud/claude-haiku-4-5-20251101" }
    }
  },
  "cron": { "maxConcurrentRuns": 4 }
}

Workaround

Moved critical monitoring jobs to macOS system crontab (running Python scripts directly, no LLM needed). Daily tasks moved to HEARTBEAT.md with manual time-based checks.