Fix #1086: XE sessions self-heal and missing sessions are surfaced (Lite + Dashboard)

[erikdarlingdata]

Fixes #1086.

Problem

The PerformanceMonitor_BlockedProcess / PerformanceMonitor_Deadlock Extended Events sessions could be absent while the blocking/deadlock collectors read the non-existent ring buffer, got zero rows, and reported SUCCESS — capture silently dead, no signal. Three variants, all confirmed:

1. Lite (the filed issue): sessions were created only on the tab-open path; the background loop never created or retried them. A failed/never-run first attempt stayed broken until a manual tab re-open.
2. Dashboard server-scoped: sessions created once at install; if later stopped/dropped, procs 22/24 swallowed the error and logged SUCCESS/0 forever.
3. Dashboard Azure SQL DB: comments claimed the database-scoped sessions were "auto-created by the collection procedures" — nothing created them anywhere. Capture was 100% non-functional on Azure SQL DB. Bonus latent bug: the Azure deadlock read filtered on xml_deadlock_report instead of database_xml_deadlock_report, so it would have returned zero events even with a session present.

Fix

Lite

• Session ensure moved into RunCollectorAsync (gated to the two XE collectors) — both tab-open and background paths now create/start/retry every cycle; existence check is one cheap round-trip once healthy.
• Ensure failures throw (XeSessionEnsureException) instead of being swallowed, classified PERMISSIONS (229/297/300, keeps existing skip-until-restart semantics) or ERROR. Ensure runs before the read, so zero-rows-SUCCESS can never mask a missing session.
• Surfacing: XeSessionUnavailable on the health entry, status-bar "Capture down" branch (PERMISSIONS failures don't increment ConsecutiveErrors and were otherwise invisible), edge-triggered tray balloon per server.
• Azure DDL gains STARTUP_STATE = ON (MS-recommended; sessions restart after failover — this was Lite's "stops on reconnect" symptom).

Dashboard

• Procs 22/24 ensure the session at the top of every run (before BEGIN TRANSACTION — XE DDL isn't allowed inside one; Azure-only catalog views accessed via dynamic SQL per the procs' existing pattern).
• Azure SQL DB database-scoped sessions are now actually created (database_xml_deadlock_report for deadlocks) + the read-side event-name fix.
• Honest logging: genuinely-absent-and-uncreatable session → SESSION_MISSING with the real error, not SUCCESS. Ring-buffer read errors re-raise.
• New Capture Down / Capture Restored alert via the standard engine (snoozable tray + email + webhook + history + cooldown + mute), fed by the latest collection_log status through GetAlertHealthAsync — no new polling loop. Gated on the blocking/deadlock notify prefs.

No upgrades/ entry: proc-body changes only (idempotent install/*), and SESSION_MISSING (15 chars) fits the existing unconstrained nvarchar(20) column. No csproj version bump — dev bumps at release time and the 2.11.0→2.12.0 train is owned by parallel work.

Known limitation

On Azure SQL DB the blocked-process threshold can't be set via sp_configure and MS documents no default — the blocked-process session may exist yet capture nothing there (deadlock capture has no such dependency). Verified against MS Learn; needs a live Azure SQL DB target to confirm runtime behavior. Called out in the proc header.

Test plan

• 5 new XeSessionHealthTests regression-lock the silent-OK → surfaced → self-heal bookkeeping (incl. the PERMISSIONS-invisible case and stale-message clearing)
• Lite suite 411/411; Dashboard suite 476/476; both apps build clean
• Live on SQL2019: procs deploy idempotently; drop both sessions → next run recreates them (running, STARTUP_STATE = ON); stop session → next run restarts it; double-run is a no-op
• Live on SQL2019: low-priv login (VIEW SERVER STATE, no ALTER ANY EVENT SESSION) + absent session → logs SESSION_MISSING with "User does not have permission…"; next privileged run logs SUCCESS and heals the session
• Azure SQL DB runtime verification (no target available)

🤖 Generated with Claude Code