Problem

PR #6591 added fingerprints to the poll-alerts Pusher payload, enabling consumers to target specific alerts instead of doing a full-table refresh. However, subscribers still cannot determine whether an alert transitioned to resolved (or any other status) without making a follow-up REST call to /alerts/query or /alerts/batch.

Since poll-alerts fires for any enrichment (new alert, ack, resolve, suppress), consumers that only care about resolution must query Keep on every single event -- even when no alert actually resolved. This adds unnecessary load on the Keep API.

Proposal

Extend the poll-alerts payload with backward-compatible, additive fields:

{
  "fingerprints": ["fp-1", "fp-2"],
  "alerts": [
    {"fingerprint": "fp-1", "status": "resolved", "previous_status": "acknowledged"},
    {"fingerprint": "fp-2", "status": "firing", "previous_status": null}
  ],
  "statuses": {"fp-1": "resolved", "fp-2": "firing"},
  "resolved_fingerprints": ["fp-1"]
}

• fingerprints -- unchanged (backward compat with feat: include fingerprints in poll-alerts payload and patch visible alert rows #6591)
• alerts -- per-alert transition with current and previous status
• statuses -- fingerprint-to-status map for quick lookup
• resolved_fingerprints -- convenience subset where status == "resolved"

Respects existing FINGERPRINT_PAYLOAD_LIMIT (100): if over limit, falls back to {"fingerprints": []} and omits transition fields.

How this helps downstream consumers

Event-driven retirement without extra API calls:

Downstream services (e.g. reconcilers, external webhooks, lifecycle managers) can now:

1. Receive poll-alerts event
2. Check resolved_fingerprints (or filter statuses for resolved)
3. Act immediately (retire proposals, close tickets, update dashboards)
4. Skip the /alerts/query round-trip entirely

This turns a poll-then-query pattern into a pure event-driven flow -- reducing Keep API load and cutting end-to-end latency for status-aware consumers from seconds to milliseconds.

Use cases enabled:

• Reconcilers that retire pending suggestions when alerts resolve
• Webhook integrations that only care about specific transitions (e.g. * -> resolved)
• Dashboards that track real-time status distribution without polling the REST API
• Audit systems that log state transitions with previous-status context

Implementation

See PR #6601 which implements this feature:

• poll_alerts_payload() helper centralizes payload construction
• Previous status captured via batch DB query before set_last_alert() overwrites it
• UI type widened but behavior unchanged (no breaking change for existing consumers)
• Comprehensive test coverage for payload builder and UI parser

Non-goals (potential follow-ups)

• Dedicated alert-resolved Pusher channel (separate subscription surface)
• Filtering poll-alerts to only fire on real transitions (currently fires on all enrichments)
• Server-side subscription filters per consumer