fix(heartbeat): keep routine ok states silent

[rcnsnr]

Summary

• Update heartbeat prompts in loop.py and cron.py to instruct agents to use notify_user only for actionable updates, alerts, or scheduled digest outputs
• When everything is normal, agents should stay silent and return HEARTBEAT_OK without notifying the user

Why

The previous prompts always instructed agents to "report results to the user via notify_user", producing unnecessary notifications on every routine heartbeat cycle where nothing meaningful happened. This creates notification fatigue and wastes API credits on redundant notify_user tool calls.

Test plan

• test_heartbeat_prompt_uses_silent_ok_rule — verifies new prompt text is present and old phrasing is gone
• Full TestHeartbeatDispatchFn suite: 5 passed, 0 failed

[cla-assistant]

All committers have signed the CLA.

[cla-assistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[bicced]

Approving — sound, low-risk improvement. Traced the full heartbeat/notification path to confirm safety:

• Silent heartbeats are already a designed-and-tested outcome: suppress_empty + _EMPTY_RESPONSES (cron.py) drop HEARTBEAT_OK, the heartbeat_complete liveness event is emitted independently of any notify_user call, and no downstream consumer expects a per-heartbeat notification.
• Daily summaries are unaffected — they run via the deterministic compose_work_summary cron path (ensure_summary_job), not the heartbeat prompt.
• The "actionable / alert / scheduled digest" carve-out preserves notifications for monitor/listening-style digest agents.
• Aligns with the chat-native-delivery direction: notify_user now writes a notification-role transcript row, so routine "nothing happened" pings are higher-visibility noise — trimming them is consistent.

One note on leverage: the heaviest no-op ticks are already short-circuited structurally before the LLM ever runs (the skip-LLM paths in cron.py / loop.py + suppress_empty), so this prompt nudge mainly trims over-reporting agents. That's fine — just don't expect a dramatic drop. If a hard guarantee is ever wanted, a server-side filter on heartbeat-origin notification events would be the structural lever (out of scope here).

Verified locally: tests/test_cron.py 92 passed, tests/test_loop.py 223 passed; the notify_user to report results assertions at test_loop.py:1999/2016 are unaffected (they cover the task/chat prompts, not the heartbeat prompt this edits).