feat: AI categorisation fallback, unsubscribe confidence scoring, one-line summaries

Heuristic classifier falls back to AI only when it can't determine a category;
unsubscribe confidence blends method reliability with an AI safety opinion and
surfaces it in the Unsubscribe Manager; emails get an on-demand AI one-line
summary in the detail pane. All AI calls degrade gracefully when AI is disabled
or the provider errors.

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

src/InboxIntel.Domain/Entities/UnsubscribeItem.cs

@@ -26,4 +26,11 @@ public class UnsubscribeItem : AuditableEntity
 
     public DateTimeOffset? LastAttemptUtc { get; set; }
     public string? ResultMessage { get; set; }
+
+    /// <summary>
+    /// 0-1 confidence that unsubscribing is safe (won't cut off something the
+    /// user actually wants, e.g. security alerts or account notices). Combines
+    /// a method/volume heuristic with an AI opinion when AI is enabled.
+    /// </summary>
+    public double Confidence { get; set; }
 }