AG alerts: post-release follow-ups (Lite parity, cross-node dedup, disconnect re-fire)

[erikdarlingdata]

The #1692 AG alert family shipped with three deliberate, stated limits. Address after the current release:

1. Lite AG alert parity — Lite collects both AG grains (Availability Group health collection in Lite + Darling (#991) #1688) but has no AG alerts. JudgeAgSync is a pure static designed to lift into PerformanceMonitor.Common (the ConnectionAlertPolicy/Connection alerts never fire for a server that is already down, and never re-fire during a standing outage #1659 discipline: shared class, identical test matrices in Lite.Tests + Darling.Tests). Wire Lite's alert path to the same four alerts with the same metric names.
2. Cross-node de-duplication — a 3-node AG with all nodes monitored reports one role change once per monitored server. The replica-grain collector doesn't carry is_local; add it (V-next migration + collector column) and de-dup fleet-side (prefer the primary's perspective, or is_local=1 rows) in the Darling evaluator.
3. AG Replica Disconnected re-fire — currently a pure edge; a standing disconnect never re-announces. Extend with the Connection alerts never fire for a server that is already down, and never re-fire during a standing outage #1659 opt-in re-fire pattern (same ConnectionAlertPolicy-style clock, stamp-on-delivery, default off).

Also from the same round, still open elsewhere: #1687 (compose HardRowCap honest truncation). Topology-page follow-ons (failover history timeline off the collected role edges) can ride with item 2.

🤖 Generated with Claude Code

[erikdarlingdata]

Complete. All three legs merged to dev:

• Lite parity — Lite gets the Availability Group alerts, off a shared policy (#1696) #1726: Lite raises the same four AG conditions under the same metric names, off a shared PerformanceMonitor.Common.AgAlertPolicy (Darling refactored onto it with no behavior change).
• Cross-node de-duplication — Report an AG failover once, and re-fire a standing replica disconnect (#1696) #1734: ag_replica_states.is_local (read from the DMV, not inferred) plus vantage ranking so one server judges each AG; the AG edge state drops the server id because an AG is one object however many nodes watch it.
• Disconnect re-fire — Report an AG failover once, and re-fire a standing replica disconnect (#1696) #1734: ag_disconnect_refire_minutes (default 0 = off, clamped 0-1440), stamped on delivery only so a suppressed alert cannot consume the window, cleared on reconnect.

Both columns ride store migration V37. The rendering half followed in #1735, which surfaces is_local on every AG surface so the vantage the dedup depends on is visible to an operator.