Recs rebuild D2: persist the built RemediationAction (P2 reorder)

[erikdarlingdata]

Scope

Implements D2 (persist the built RemediationAction) + P2 (pipeline reorder) of the recommendations engine rebuild. Dashboard-only. Lite is unchanged (advise/copy-paste only; it renders advice from the fact-key and needs no persisted action). Does NOT build the Recommendations UI (WS1) or new config facts (WS3).

The (later) Recommendations surface must show Apply + the two-sided consent gate for findings read back from storage. The remediation builders REQUIRE finding.DrillDown, but GetRecentFindingsAsync returns findings with DrillDown == null. So — exactly as the existing Alert path does with ContextJson — we persist the BUILT RemediationAction on the finding row and deserialize it on read. (D7, already merged, ensures config/RCSI findings carry their config_issues drill-down below severity 0.5 so their actions can be built.)

File:line changelog

• PerformanceMonitor.Notifications/AlertContext.cs (+36): added public AlertContextSerializer.SerializeAction(RemediationAction?) (:169) and DeserializeAction(string?) (:184) that wrap the EXISTING private ToDto/FromDto (the proven round-trip for RcsiInactionFigures/ClearPlanFigures/all target lists). DeserializeAction is null/try-catch on bad json, mirroring TryDeserialize.
• PerformanceMonitor.Analysis/AnalysisModels.cs (+14): AnalysisFinding.Remediation (:126) — ephemeral like DrillDown (built post-enrich on write, deserialized on read; not a scored field). RemediationAction is in the same assembly.
• Dashboard/Analysis/SqlServerFindingStore.cs (net +63):
  • remediation_action_json nvarchar(max) NULL added to the canonical CREATE TABLE (:59) AND an idempotent IF COL_LENGTH(...) IS NULL ALTER TABLE ... ADD migration in EnsureTablesExistAsync (:88-92). App-managed schema — nothing added to install/ or upgrades/.
  • Split SaveFindingsAsync → FilterMutedFindingsAsync (:106, mute-filter survivors, NO insert) + InsertFindingsAsync (:171, batched insert persisting the action json). Mute/dedup/ordering + _nextId sequencing identical; both keep PR-1's single-connection + single-EnsureTablesExistAsync discipline.
  • InsertFindingAsync (:404-438): added the column/param; persists AlertContextSerializer.SerializeAction(finding.Remediation).
  • GetRecentFindingsAsync SELECT (:216) + ReadFinding (:451): selects/deserializes ordinal 18, field-count-guarded (reader.FieldCount > 18) so GetLatestFindingsAsync (which does not select it) is unaffected.
• Dashboard/Analysis/AnalysisService.cs (net +35): AnalyzeAsync reorder (:145-175) — filter survivors → enrich survivors → build+attach RemediationAction (try BuildAction ?? BuildRcsiAction ?? BuildClearPlanAction) → batched insert. The returned/enriched (now action-bearing) list still flows to AnalysisCompleted + (via the scheduler) NotifyAsync unchanged.
• Dashboard/Services/AnalysisScheduler.cs (+1/-1): stale SaveFindingsAsync comment reference updated (no behavior change; notify path verified intact at :193-198).
• Dashboard.Tests/AnalysisNotificationTests.cs (+93): 4 new tests (below).

Tests (real results)

• dotnet build PerformanceMonitor.sln -c Debug → 0 errors (1 pre-existing unrelated CS0649 in RemediationTests.cs).
• Dashboard.Tests: 238 passed / 0 failed / 0 skipped.
• Lite.Tests: 360 passed / 0 failed / 0 skipped.
• New tests (all green, confirmed present via --list-tests):
  • SerializeAction_RcsiAction_RoundTripsFiguresAndTargets — RCSI action through SerializeAction/DeserializeAction; asserts RcsiInactionFigures (BlockingEvents/Deadlocks/ReaderWriterPct) survive.
  • SerializeAction_ClearPlanAction_RoundTripsFiguresAndTargets — clear-plan action; asserts ClearPlanFigures (CpuPercent/AnomalyRatio) survive.
  • SerializeAction_NullAction_SerializesAndDeserializesToNull — null/empty/garbage json → null, no throw.
  • PersistedRcsiAction_RendersTwoSidedConsentGate_WithFindingNull (the C1 gating test) — synth RCSI finding (rcsi:false + inaction enrichment) → serialize → deserialize → FactRiskDisclosure.GetForAction(action, finding: null) returns a non-null two-sided RiskDisclosure with BOTH RisksOfChanging and RisksOfNotChanging non-empty AND the original inaction figures present ("12 blocked-process events", "3 deadlocks", "80%", "RCSI eliminates"). Proves the consent gate renders from the persisted action alone — exactly how the real Apply surface calls it.

Needs integration verification (not unit-testable)

config.analysis_findings is SQL Server; there is no in-memory harness for the store, so the following require a real DB and are NOT claimed as covered by unit tests:

1. Full persist → reload round-trip: run AnalyzeAsync against a server with an RCSI-off / config finding; confirm a row is written with non-null remediation_action_json, then GetRecentFindingsAsync reads it back with finding.Remediation != null and the consent gate renders.
2. Idempotent migration on a pre-existing DB: against a config.analysis_findings created before this change, confirm EnsureTablesExistAsync adds the column once and is a no-op on subsequent runs (and that fresh-create DBs already have it, so the ALTER is skipped).
3. Notify path end-to-end with AnalysisNotificationsEnabled on: confirm NotifyAsync still receives the enriched findings after the reorder.
4. Cost at the default-on 30-min interval across N servers (one extra connection-open per cycle from the filter/insert split).

🤖 Generated with Claude Code