feat: expand AI moderation with structured analysis, review workflow, and guardrails

- Add structured AI moderation fields (categories, severity, confidence,
  recommended_action, policy_version, evidence) to messages table
- Add moderation_reviews, moderation_actions, and retention_policies tables
- Upgrade LLM response parsing to support structured metadata with backwards
  compatibility for legacy responses
- Implement public AI evaluation review UI with decision controls
  (approve, false positive + reanalyze, escalate)
- Add auto-delete guardrails requiring high confidence, severity, and
  allowed categories; log all attempts to moderation_actions
- Add retention manager scaffolding for messages/attachments/voice
- Add action executor for moderation actions (mute, warn, kick, ban)
- Add review routes: GET/POST/PATCH /api/reviews, GET/POST/PATCH /api/actions
- Preserve auth separation: voice/media/recordings gated, review public

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

src/moderation/aiAnalyzer.ts

@@ -307,6 +307,12 @@ async function processIndividualFallback(
         score: r.score,
         raw: JSON.stringify(analysisResult.raw),
         analysis: r.analysis,
+        categories: r.categories,
+        severity: r.severity,
+        confidence: r.confidence,
+        recommendedAction: r.recommendedAction,
+        policyVersion: r.policyVersion,
+        evidence: r.evidence,
         analyzedAt: Date.now(),
         error: null,
       },
@@ -359,6 +365,12 @@ async function processIndividualFallback(
             raw: null,
             analysis:
               "Individual fallback exhausted all retries: LLM consistently dropped this message even in single-target mode",
+            categories: ["individual_analysis_exhausted"],
+            severity: "none",
+            confidence: 0,
+            recommendedAction: "review",
+            policyVersion: "default-2026-05-30",
+            evidence: [],
             analyzedAt: Date.now(),
             error: lastError,
           },