Separate alerts that should page from noise

Audit these alerts.

For each, answer one question first: **what does a human do, right now, when this fires?**
- There is an immediate action → page
- Action needed but not urgent → ticket
- No action, only information → dashboard, not an alert
- *Fires regularly and is routinely ignored → delete it.* An ignored alert trains people to ignore the next one.

Then classify each and give:
- Threshold and duration. Most noisy alerts fire on an instantaneous value that should have a sustained window.
- Whether it alerts on a **symptom users feel** or on a cause. Symptom alerts catch problems you did not predict; cause alerts mostly fire without user impact.
- What it would miss — the failure mode it does not cover.

Then:
1. Alerts to delete, with why. Be specific; this is the highest-value output.
2. **Missing alerts** — from the recent incidents I gave you, which ones nothing would have caught. That gap matters more than any tuning.
3. Alerts that fire together, which should be one alert.
4. The night-time page list, kept as short as it can be.

Rules:
- *Deleting alerts is the main work here.* Do not hedge by keeping everything at a lower severity.
- An alert with no runbook is an alert nobody can act on. Note which lack one.